Our corrections policy
Some of what we publish will turn out to be wrong. This page sets out how to tell us, how quickly we fix each class of error, and what we do when a reported mistake cannot be confirmed. It applies to every review, ranking, and comparison on this site.
Why this page exists
AI tools change their pricing, their limits, and their feature sets constantly. A plan that was accurate on the day we published can be wrong a month later, and a claim can be wrong the moment it goes live because a vendor updated a page while we were writing about it.
We hold vendors to their own documentation. The same standard has to apply to us, which means fixing our errors in public, with a date, instead of editing them away quietly. Accuracy beats being right the first time.
Correction levels and response times
We sort corrections into four levels. The level decides how fast we act and how visible the fix is. For material and verdict-affecting errors, target windows run from the point we confirm the error, not from the moment a report reaches us.
| Level | Covers | Target response | Shown as |
|---|---|---|---|
| S1: Minor | Typos, broken links, formatting | Within 72 hours | Silent fix |
| S2: Material | Price, feature, date, availability errors | Within 24 hours of confirmation | Dated correction note on the page |
| S3: Verdict-affecting | Facts that change a score or recommendation | Same day as confirmation | Prominent note + updated verdict |
| S4: Ethical/legal | Misattribution, conflict-of-interest, data issues | Immediate review | Editor's note; retraction if warranted |
A silent fix at S1 means we correct the page without a note, because nothing about the substance changed. Everything from S2 upward is dated and stated on the page itself.
How to request a correction
Email info@aitoolspolice.com. Anyone can report an error: readers, vendors, and people who work on a competing product. Every report is checked the same way, whoever sends it.
Include:
- — The page URL, and where on the page the claim sits.
- — The claim as we published it, quoted.
- — The correct information, with a source we can open: a live pricing page, official documentation, a policy or legal page, or a public filing.
A report without a checkable source still gets read, but it takes longer to resolve, and we do not change a published claim on assertion alone.
What happens next
We check the reported claim against primary sources first. That means the vendor's live pricing page, the official documentation, or the relevant legal text. Independent user reports often tell us where to look, but they do not on their own establish what a plan costs or what a feature does.
If the error is confirmed, we fix it inside the target window for its level and date the fix on the page. From S2 upward, the note says what was wrong and what the correct information is, so you can see what changed instead of trusting that something did.
If we cannot confirm the error, the original stands. Where the disagreement is material, we note on the page that the claim is disputed and say who disputes it. We would rather publish a labeled disagreement than delete a finding or present a contested claim as settled. Either way, we reply to whoever reported it.
The final call on a correction is a human editorial decision. Mucahit Kaya founded and edits AI Tools Police and is accountable for what stays on the page.
What vendors can and cannot do
Vendors are welcome to report factual errors in our coverage of their product, and those reports get exactly the same treatment as a reader's.
What vendors can do
Flag a factual error at any time after publication, with a source we can check. That includes a price change, a renamed or retired plan, a deprecated feature, a new usage limit, or a policy update we have not caught yet. If a vendor believes we have read their documentation wrong, they can say so and point at the passage.
What vendors cannot do
No vendor sees a draft before it is published. No vendor edits our wording or signs off on a verdict. This holds for free coverage and for paid submissions equally: an expedited investigation buys a place at the front of the queue, a sponsored placement buys a slot that is labeled as sponsored wherever it appears, and neither buys a score or a ranking position.
A correction changes a verdict only when the underlying evidence changes. If a price or a limit we published turns out to be wrong, the assessment moves with the corrected number. Disagreement with our conclusion, unaccompanied by new evidence, does not.
Update discipline
We record the date each review was first published, and that date never changes. The date shown on a page is its last-updated date, recorded separately from the original publication date.
The last-updated date moves only when something material changed on the page: a price, a usage limit, a feature, a verdict, a ranking position. Copy-editing and formatting do not reset it. A fresh date on a page whose facts have not been rechecked is a false signal, and we treat it as one.
Material updates are dated on the page itself. That way you can see when the facts were last checked, rather than inferring it from the design.
Corrections log
Every material correction we have made, newest first. Each entry says what we published, what is right, and how the error came to light, and links to the page that carried the error. Where the fix was made on the page, that page carries the dated note; where it was made in the site's template or release process, this log is the note. Minor fixes are silent by the policy above and are not listed. The log begins with our first dated correction, in July 2026.
- What we published.
- From 29 August 2026 this page carried a dated correction withdrawing the figure that the $10 Basic plan buys "about 200 fast images", explaining that Basic's 200 fast GPU minutes had been misread as images. Two sentences kept the withdrawn figure anyway. The pricing paragraph still glossed Basic's 3.3 fast GPU hours as "roughly 200 images", and the closing verdict still told readers to skip the Basic plan "since a reported ~200 images is not a working month".
- What is right.
- A number this review had already published as false was doing the work of a buying recommendation, which is the part worth naming. Both sentences are rewritten around what Midjourney actually publishes: 200 fast GPU minutes for Basic, which is 3.3 fast GPU hours, and no images-per-plan figure of any kind. Where the page now gives a sense of scale it is labelled as our own arithmetic from the vendor's documented rate, not as a vendor number. The verdict no longer tells anyone to skip Basic on volume grounds, because the withdrawn figure had understated what $10 buys rather than overstating it. What the page says about Basic now rests on two documented limits: no Relax Mode, so generation stops when the fast minutes are spent, and no Stealth Mode, so every image lands in the public gallery. Basic is described as a paid evaluation of the model rather than a plan to skip. The 4.2 rating is unchanged. A correction that lands on part of a page and not the rest leaves the claim in circulation, and this site now checks for that case directly.
- How it was found.
- A sweep of our own live pages, when a writer cleaning punctuation on the file noticed that the page contradicted its own correction.
- What we published.
- Since 12 July 2026, that G2 showed 4.5 out of 5 for Julius AI from four reviews, all dated 2024 and invite-sourced, and that "G2's own page states there is not enough data to provide buying insight". No read date and no link travelled with any of it.
- What is right.
- Read at g2.com/products/julius-ai-julius/reviews on 9 September 2026, that listing shows 4.5 out of 5 from 4 reviews, so the figure itself stands and now carries its count, its read date and a link. Two things around it did not. The claim about a G2 notice is withdrawn: no such notice is on the listing as read that day, and we do not say when or why it changed. The descriptions "dated 2024" and "invite-sourced" are dropped rather than re-dated, because our read did not cover them and a read date must not appear to vouch for what it did not check. The page also gains a fact a reader needs in order to check us: G2 carries two listings for this product, and the other one, titled Julius AI at g2.com/products/julius-ai/reviews, is an unclaimed profile showing no reviews at all, so anyone verifying the 4.5 can land on either page and reach the opposite conclusion. The existing point that one of the four reviews is about the unrelated influencer-marketing Julius is now specific: it is the review titled "Best in segment activity dashboard and Campaigning". Four reviews cannot be weighed against the App Store's 399, and the count remains the finding. The 3.7 rating is unchanged.
- How it was found.
- A sweep of our own live pages against our own sourcing standard. The first attempt at this correction was wrong, because the unclaimed listing was read as though it were the product's only one, and a writer's check of the review's own structured data caught it before it published.
- What we published.
- Since 2 September 2026, on ten surfaces including a section heading, a Trustpilot rating of 2.7 out of 5 across 81 reviews with "roughly 60% of those reviews negative". The figure carried no read date and no link, and the share of negative reviews was ours rather than the platform's.
- What is right.
- Read at trustpilot.com/review/krea.ai on 9 September 2026, Krea AI is 3.6 out of 5 across 149 reviews. Because the old figure carried no read date, we cannot say when it was accurate, only that it is not accurate now. The negative share was wrong on its own terms as well: Trustpilot's published breakdown of those 149 reviews puts one star and two star together at 44%, not roughly 60%. That breakdown now stands in place of a single number, because the shape is the useful part: 47% five star and 35% one star, with 6%, 3% and 9% across the middle. Four facts about the profile are printed beside each other and deliberately not joined into a chain: the score is 3.6 across 149; Krea claimed the profile in June 2026; the profile carries a paid Trustpilot subscription; and Trustpilot tags reviews on it as invited. Whether the invitations moved the score is not something we can establish from outside the company, so it is left unresolved rather than asserted. The small sample argument stays, with the number corrected, but the clause saying Trustpilot skews toward people motivated to complain comes out, because it is not true of a pool the company invites. The 3.5 rating is unchanged.
- How it was found.
- A sweep of our own live pages against our own sourcing standard, three days after the same standard was applied to eighteen other reviews. This file was not among them.
- What we published.
- Since 24 July 2026, in a pros bullet, an FAQ answer and the opening section, "a Trustpilot score cited around 4.9 out of 5 from more than 5,800 reviews". The page also stated, in its sourcing note, that ratings on G2 and Trustpilot were taken from indexed summaries rather than read directly because both sites block automated access.
- What is right.
- Read at trustpilot.com/review/aragon.ai on 9 September 2026, Aragon AI is 4.7 out of 5 across 6,852 reviews. The hedge in "cited around" was the tell that nobody had opened the profile, and the sourcing note said as much in plain words, which made the claim and its own disclaimer the same defect twice. What the profile also shows belongs beside the number rather than in a footnote, because this page used the score as evidence that Aragon is large and established: Trustpilot states the profile was claimed in April 2024, that it carries a paid Trustpilot subscription, and that the individual reviews are tagged invited, which means the pool is drawn from customers Aragon asked. Trustpilot's published breakdown of the 6,852 runs 86% five star, 11% four, 1% three, under 1% two and 1% one. The score stands as evidence of scale; the sixteen dated user reports lower on the page carry the questions an invited pool is least likely to answer. The sourcing note now says what is true: scripted retrieval of these platforms fails, so the figures are read in a browser and dated. The 3.7 rating is unchanged.
- How it was found.
- The same sweep. The self-contradiction between the figure and the sourcing note was visible on the page the whole time.
- What we published.
- Since 2 September 2026: "On Trustpilot, Midjourney reportedly carries a strikingly low rating, around 1.5 stars." No count, no read date, no link, and the word reportedly standing in for a source.
- What is right.
- Read at trustpilot.com/review/midjourney.com on 9 September 2026, Midjourney is 1.6 out of 5 across 372 reviews. Trustpilot's published breakdown of those reviews is 77% one star against 10% five star, with 4%, 4% and 5% between. Two facts about the pool now travel with the number. Trustpilot prints on the profile, claimed in June 2023, that the company has no recent history of asking for reviews, so nobody is soliciting these ratings and everyone in the pool went looking for somewhere to post. And 372 reviews describes the people who chose to write one, not the user base. The section's existing argument is untouched, that the complaints are about billing, refunds and support rather than image quality; it now rests on a figure a reader can open. The 4.2 rating is unchanged.
- How it was found.
- The same sweep.
- What we published.
- Since 20 June 2026: "On Product Hunt, the same product sits near 1.5 out of 5", with no count and no read date, set directly against a Trustpilot score of 4.2 across 56,071 reviews as the two halves of a polarised reputation.
- What is right.
- Read at producthunt.com/products/resume-io/reviews on 9 September 2026, Product Hunt is 1.3 out of 5 from 38 reviews. The count is the correction. Thirty eight ratings and fifty six thousand ratings are not two comparable scoreboards, and the section was built as though they were: one is a sample so large that no single opinion moves it, the other is small enough that a dozen annoyed reviewers set the number outright. The audience explanation stays, because it does explain the direction of the gap, but it can no longer carry the section, since a difference in audience does not make two samples equal weights in a verdict. The Trustpilot figure was already correctly sourced and is unchanged. The 4.0 rating is unchanged.
- How it was found.
- The same sweep. The defect is the one the Gamma review carried in July, eight ratings weighed against a hundred and twenty two, appearing in a second place.
- What we published.
- Since 23 June 2026: "professional reviewers routinely score Wix around 4.9 out of 5, while real-user ratings on platforms like Trustpilot can sit far lower, in the 1.7 to 3.2 range", followed by a paragraph explaining that gap. No publication was named for the 4.9, no platform for the range, and no figure carried a count, a date or a link.
- What is right.
- The claim about what professional reviewers score is withdrawn rather than sourced, because no publication could be named for it and inventing one was the only alternative. In its place the section reports what can be opened: read at trustpilot.com/review/wix.com on 9 September 2026, Wix is 3.5 out of 5 across 28,939 reviews, with Trustpilot's published breakdown at 71% five star and 22% one star and almost nothing in between. Trustpilot also states the profile was claimed in May 2014, carries a paid Trustpilot subscription, and tags its reviews as invited. The two humped split is what the rest of the page already describes: how fast Wix gets a site live at one end, and billing disputes, support friction and lock in at the other. The body of the page also carried a hand typed "last updated June 23, 2026" while the page header printed the real modification date, so a reader could see two different dates for the same page; that line is gone and the date now comes from one place. The 4.2 rating is unchanged.
- How it was found.
- The same sweep, plus a check of every page that types a date into its own body instead of taking the one the page already renders.
- What we published.
- Since 23 June 2026: "top10.com's review cites about 84% Excellent across roughly 139,000 reviews". A Trustpilot distribution attributed to a commercial comparison site rather than to Trustpilot, with no read date.
- What is right.
- Read at trustpilot.com/review/godaddy.com on 9 September 2026, GoDaddy is 4.4 out of 5 across 142,623 reviews, and Trustpilot's own published breakdown is 84% five star, 4% four, 2% three, 1% two and 9% one. The 84% survives, as Trustpilot's own five star share rather than as another site's summary of it. Citing a comparison site for a figure we can read at its source was the defect, and the undated count had already moved. The page separately said there was a gap worth explaining rather than burying, between GoDaddy's high Trustpilot score and how critical the hosting and small business communities are. Trustpilot answers that itself: the profile has been claimed since May 2014, carries a paid Trustpilot subscription, and tags individual reviews as invited, which is Trustpilot's own label for a review left after the business asked. Soliciting reviews that way is permitted and disclosed, and it means the two pools are not assembled alike. The 3.4 rating is unchanged.
- How it was found.
- The same sweep. This one had been carried as known debt since June and was never closed.
- What we published.
- On four reviews, in various forms since June 2026, a sourcing note stating that G2, Trustpilot and Capterra block automated retrieval and that their ratings are therefore "reported as they appear in current published coverage rather than pulled directly". On the Synthesia review the same sentence then gave a figure with our own read date attached, which cannot both be true.
- What is right.
- The claim overstated what these platforms do. Scripted retrieval of them fails, and an automated fetch returns nothing usable; the pages themselves open in an ordinary browser, which is how we read them, and that is what the notes now say. Where a figure carries a read date it was read that way and the date is the date it was read. No figure was changed by this correction on its own, but it is what the two rating corrections dated today rest on, because a page that tells a reader we could not open a source has no business publishing a number from it. Where a figure still has no read date, none has been invented.
- How it was found.
- Reading both platforms in a browser while checking a different page's ratings, which made the disclaimer visibly untrue.
- What we published.
- Since 6 June 2026, a Trustpilot score of 2.3 out of 5 and a G2 score of 4.9 out of 5, neither carrying a review count, a read date or a link. The difference between them was described as "a 2.6-point gap that every other Vidnoz review we found either ignores or mentions in passing without explaining", and a dedicated section, a key takeaway, an FAQ answer and the closing verdict were all built on explaining it.
- What is right.
- Read at g2.com/products/vidnoz-ai/reviews on 9 September 2026, G2 is 4.9 out of 5 across 16 reviews. Read at trustpilot.com/review/vidnoz.com the same day, Trustpilot is 2.1 out of 5 across 83 reviews. The published 2.3 was wrong and the real difference is 2.8, but the missing count was the larger defect. Sixteen reviews cannot characterise a product and cannot be set against a pool five times its size, so the section no longer treats the two scores as comparable measurements whose difference needs explaining away. The two-populations argument survives as a secondary point rather than as the thing the section is about. In its place the page reports what Trustpilot publishes about those 83 reviews: 46% one star against 36% five star, with 18% across the three middle bands, which puts the split inside a single platform rather than between two. Nothing is derived from either score. Trustpilot's notice that the company has not invited customers recently travels with the 2.1, because who ends up in a review pool is a question about composition. The 4.0 rating is unchanged. A link in the review's structured data also pointed at a G2 profile that does not exist, and now points at the one that was read.
- How it was found.
- A sweep of our own live pages against our own sourcing standard, three days after the same standard was applied to eighteen other reviews. This file was not among them.
- What we published.
- Published on 6 September 2026, in the Kling AI entry in this log and in the body of that review: that a score of 1.2 out of 5 across 380 reviews bounds the distribution, so that even if every review that is not one star were two stars, four in five would still have to be one star. The reasoning was ours rather than a source's.
- What is right.
- The algebra holds for an arithmetic mean, and a TrustScore is not one. Trustpilot's help page on how the score is calculated, read on 6 September 2026, states that a TrustScore "isn't just a simple average of all the reviews on the profile", that newer reviews hold more weight in it than older ones, and that "our weighted average includes 7 neutral (3.5★) reviews to the calculation". So nothing about how those 380 reviews split follows from the published figure, in either direction, and we never read the star histogram. The bound comes out of the review body and out of the entry that carried it, and it is not replaced with a hedged version of itself, because a bound with a qualifier attached is the same inference. What replaces it is the limit: the score is 1.2 out of 5 across 380 reviews, read at its source on 6 September 2026 and published there as a weighted average, and it tells us nothing about the shape of what is behind it. The score did not change, and neither did the 3.7 rating. Two things that never rested on the arithmetic stay on the page. Trustpilot states that Kling has invited customers recently, which is about who ends up in the review pool rather than how the pool is weighted, and the same help page says the calculation "doesn't distinguish between verified, invited, or unprompted reviews". And the narrower claim survives, that the complaints are about the account rather than the model. The standing rule from here is that no page may derive a distribution, a bound or a share from a published aggregate score; where we want to say something about the shape of a review pool we read the histogram and cite it.
- How it was found.
- Reading Trustpilot's own documentation of how a TrustScore is calculated, hours after the Kling correction went live, while patching three other pages in the same sweep.
- What we published.
- From 17 July 2026, a Trustpilot score of around 1.9 to 2.0 and a G2 score of around 4.0, both undated and neither carrying a review count or a link, beside a Microsoft Store rating of about 4.3 and a Capterra rating in the high 3s. A cons bullet, two FAQ answers and a section written to explain the gap all rested on them.
- What is right.
- Read at trustpilot.com/review/gamma.app on 6 September 2026 and confirmed against that page's own rating markup, Gamma is 1.7 out of 5 across 122 reviews. Read at g2.com/products/gamma-gamma/reviews the same day, it is 4.6 out of 5 across 8 reviews. The count is the real correction. The page ran a dedicated FAQ answer asking why the Trustpilot score is so low when the G2 score is high, which set a software-review site against a complaints site as though both rested on comparable evidence. Eight ratings cannot be weighed against a hundred and twenty two, so that symmetry is withdrawn rather than flipped into the opposite claim, because the opposite claim is not true either: the product does review well and the company does frustrate people over billing and support, and the page now carries that split on its dated user reports rather than on two star averages. Trustpilot's own notice travels with the 1.7, that the profile is unclaimed and the company has not invited its customers, so the reviews may not be representative. Two figures come off the page entirely. The Microsoft Store rating was never read, because apps.microsoft.com served only a cookie layer and the listing never rendered. The Capterra figure was never re-read at all, and it carried no count, no date and no link, so it goes for the same reason in a weaker form. Neither is replaced with an approximation. The 3.9 rating is unchanged, because nothing about the product changed, only what we can source about it.
- How it was found.
- The last three files in the same sweep of our own pages against our own sourcing standard.
- What we published.
- Since 7 July 2026, a Capterra rating of 4.4 out of 5, latterly with a count of 31 reviews but never a date or a link, used in the quick verdict, the sourcing note and the user-evidence section as half of the positive case against Framer's low Trustpilot score. The G2 figure beside it, 4.5 out of 5, carried no count and no link either.
- What is right.
- The Capterra figure is withdrawn rather than restated. That page returns a bot check that does not clear on its own, we do not attempt to defeat bot checks, and a rating we cannot read at its source does not belong on the page. Nothing replaces it. The G2 figure was read at g2.com/products/framer/reviews on 6 September 2026 and holds at 4.5 out of 5, now with the thing that was missing from it: 144 reviews. That matters more here than the score. Framer's 1.6 on Trustpilot rests on 140 reviews, so the page is no longer setting a sample of unknown size against a known one, and two samples of almost identical size reaching opposite verdicts is what makes the split worth explaining rather than averaging away. The Trustpilot figure and its link, corrected earlier the same day, are unchanged, and so is the 4.0 rating.
- How it was found.
- The same sweep.
- What we published.
- Since 8 July 2026, a Capterra rating of about 4.6 out of 5 across more than 1,200 verified reviews, with no date and no link, on three surfaces: the sourcing note, the worth-it FAQ answer and the closing verdict.
- What is right.
- Withdrawn rather than restated, for the reason the Framer figure was: Capterra returns a bot check that does not clear on its own, and we do not publish a rating we cannot read at its source. Nothing replaces it. On two of the three surfaces the number was carrying an argument rather than sitting decoratively, so those sentences were rewritten instead of having the figure cut out of them. The worth-it answer and the verdict now rest on TechRadar's score and on the page's own dated user reports, including the negative half of that record, the May 2025 relaunch and the legacy app that was remotely disabled. TechRadar's 4.5 stays, with a distinction it did not carry before: it is one publication's own review score rather than an average of user ratings, which is why it was never part of this sweep. The sourcing note no longer claims rating aggregators as part of the basis, because no aggregator figure now appears on the page. The 4.1 rating and the 5 September pricing check are unchanged.
- How it was found.
- The same sweep.
- What we published.
- From 20 June 2026, a Trustpilot score of 1.1 out of 5 from 447 reviews, carried in the meta description, a cons bullet, an FAQ answer, the verdict and a section built to explain it.
- What is right.
- Read at trustpilot.com/review/www.fotor.com on 6 September 2026, the score is 3.7 out of 5 across 1,568 reviews, confirmed against the page's own rating markup. That is not a number swap. A section titled to explain a strikingly low score no longer had a low score to explain, so it was rebuilt around what the record actually shows and around the finding that never depended on the rating at all: Fotor's terms state subscription fees are non-refundable and cancellation must be requested at least 24 hours before renewal. The complaints we described are still there and still about billing; what was wrong was their weight. Trustpilot also notes the company has not recently invited reviews, which skews a sample toward complainants, and that Fotor replies to 37% of negative ones. The meta description, which was still serving the withdrawn figure to search results, now names the terms instead. The 3.8 rating is unchanged and that is a judgement worth stating: part of it was set against a billing reputation that no longer reads as bad, so if it moves it should move upward, and we would rather leave it than raise a score on a correction that runs in the vendor's favour.
- How it was found.
- A sweep of our own pages against our own sourcing standard.
- What we published.
- From 11 June 2026, a Trustpilot score of 2.8 out of 5, with no review count and no date, used in the cons, an FAQ answer, the sourcing note and the verdict.
- What is right.
- Read at trustpilot.com/review/klingai.com on 6 September 2026 it is 1.2 out of 5 across 380 reviews. Our figure was more than twice as favourable as the vendor's actual standing. The reasoning built on it is withdrawn rather than restated: the page argued that Trustpilot self-selects for grievance and that the low score therefore reflected billing friction rather than the product. What weighs against that at 1.2 is composition. Trustpilot says Kling has invited customers recently, and invitations reach the satisfied subscriber who would never otherwise post, so the mechanism that normally lifts an average was running. We did not read the star histogram, and the page now says both that and why it matters: Trustpilot publishes the TrustScore as a weighted average rather than a plain mean, so the shape of those 380 reviews cannot be read off the 1.2. A second reason first given in this entry, that the score bounds the distribution, is itself withdrawn in a separate entry of the same date. What survives is the narrower claim, that the complaints are about the account rather than the model, and it is now framed as two records measuring different things rather than one discounting the other. The pika.md comparison table, which printed the old 2.8 as a competitor figure, is corrected in the same pass.
- How it was found.
- The same sweep, extended to every page that cites these scores.
- What we published.
- A Trustpilot, G2 or Capterra star rating with no date, no review count and no link to the source, on the Beautiful.ai, ElevenLabs, Framer, GPTZero, Hailuo AI, HeyGen, LOVO AI, MyPerfectResume, Pika, Plus AI, Resume.io, Speechify and Synthesia reviews.
- What is right.
- Every one was re-read at its source in a browser on 6 September 2026 and confirmed against the page's own rating markup. Nine had moved: Beautiful.ai 2.9 to 2.8, ElevenLabs 3.1 to 3.0, Framer 1.7 to 1.6, GPTZero 2.4 to 2.2, LOVO AI 2.3 to 1.6, Pika 1.9 to 1.7, Plus AI 2.3 to 2.1, Resume.io 4.4 to 4.2, Speechify 4.6 up to 4.7. Review counts were stale on several more, by 428 on Synthesia and 466 on HeyGen. Each figure now carries its count, the date we read it and a link. The drift is not really the point: our own standard has required a date, a source and a link on every user claim since 19 August 2026, and an undated star average implies a currency it does not have. Two of these ratings moved far enough to change what the page says, and they are logged separately. One deserves naming here: Plus AI's 2.1 comes from ten reviews, which is too small a sample to characterise a product, so the count now travels with it.
- How it was found.
- A sweep of our own pages against our own sourcing standard.
- What we published.
- Since 7 June 2026, "Trustpilot user rating sits at a low 2.4/5", carried in three places with no date, no review count and no link.
- What is right.
- Read at trustpilot.com/review/humbot.ai in a browser on 5 September 2026, the score is 2.2 across 100 reviews, confirmed against that page's own rating markup. Two things Trustpilot publishes beside the number are now carried with it, because the number alone reads as a verdict it does not support: the company has not invited customers recently, so the sample skews toward people who arrived to complain, and Humbot has replied to 84% of negative reviews. The recent complaints concentrate on recurring charges and cancellation rather than output quality, which corroborates this page's own refund section. Our standard has required a date, a source and a link on every user claim since 19 August 2026; this figure predated it and was never brought up to it, the same gap the WriteHuman page was corrected for.
- How it was found.
- Closing the open list after the design and data-analysis batch shipped.
- What we published.
- In the AutoShorts.ai entry of 31 July 2026, that "the headline rating of 3.8 also never matched the page's own five axis scores, which averaged 3.6", listed among that page's errors and corrected to "the rating now reads 3.6, matching the scorecard".
- What is right.
- That treats a gap between the headline and the mean of the subscores as an error. Our methodology page says the opposite, and said so at the time: the headline score is not the average of the dimension scores, and a reader should not expect the arithmetic to work, because the dimensions are separate judgements about separate things. So this log taught readers a rule the methodology page denies, on two pages that exist to be trusted about exactly this. The standing rule is the methodology page's. A subscore moves on evidence about that dimension, the headline moves on evidence about the tool, and neither is computed from the other; where a page's headline and its axes diverge far enough to look like a mistake, the page should say why rather than have the number quietly adjusted. AutoShorts.ai stays at 3.6: the free-tier and launch-date findings in that same entry are unaffected and the score is defensible on them. What is withdrawn is the arithmetic as a reason.
- How it was found.
- An editorial review of our own scoring practice, prompted by three pages in one week holding a headline that the axes do not average to.
- What we published.
- Since 12 July 2026, Picsart's Ultra tier at "about $18 to $25 a month", with no indication that it is priced per seat.
- What is right.
- An archived capture of picsart.com/pricing dated 26 August 2026, which renders in US dollars where the live page serves Turkish lira from our location, prices Ultra at $47 a month month-to-month or $37.50 a month billed yearly, per seat. That is roughly double what we published, and a three-person team is $1,350 a year rather than the figure our page implied. Pro is restated as two prices for two commitments, $15 month-to-month and $10.50 billed yearly, and both ladders reconcile against Picsart's own "You save $54" and "You save $114" lines. A Plus tier we had listed from third-party trackers does not appear in the capture; we do not claim it was withdrawn, because one capture cannot show that. The page carries no verified-pricing stamp, because an archive is not a live read.
- How it was found.
- A pre-release check of the AI design cluster, reading every vendor pricing page before shipping the batch.
- What we published.
- Since 8 July 2026, Canva Pro at "about $15/mo" and Canva Business at "about $20/user/mo", with no billing basis stated, and a note that most sources converge on $20 while one cites $25.
- What is right.
- Those are the yearly-billing rates. Read on canva.com's US pricing view on 5 September 2026, Pro is $18 a month month-to-month or $180 a year, and Business is $25 a month per person or $250 a year per person. A reader paying monthly paid about 20% more than our page said on Pro and 25% more on Business. The outlier we dismissed was Canva's own month-to-month price, and the disagreement we reported was between third-party trackers, not about the vendor's page. Both bases are now given on this review and on the six other pages that quoted the figure while comparing another tool to it: the AI design ranking, and the Adobe Express, Picsart, Gamma, Decktopus, Beautiful.ai and Plus AI reviews.
- How it was found.
- The same pre-release check.
- What we published.
- An AI Pass add-on priced at "about $100/person/mo" and advertised at roughly 40x more AI than Pro, and monthly allowances of 50 credits on Free and 500 on Pro.
- What is right.
- Canva's pricing page no longer describes credits. It states an AI allowance in uses: up to 200 Standard AI uses or 20 Premium AI uses on Free, and paid tiers only as multiples of that, 10x on Pro and 20x on Business, drawn from one shared pool across Standard, Premium and Ultra AI. The AI Pass carries no price on that page at all; it appears only as an add-on "available (extra cost)". The $100 figure and the 40x multiple are withdrawn rather than restated, and not replaced with an approximation, because a rounded guess still lands in a reader's budget as a fact.
- How it was found.
- The same pre-release check.
- What we published.
- Since 8 July 2026, a Starter tier at $16 a month with 500 credits, and a free plan of 30 credits a month described as permanent rather than a trial.
- What is right.
- findanomaly.ai/pricing read on 5 September 2026 shows no Starter tier, and its free plan gives 15 credits, half what we said. So the cheapest paid plan we can point to is Pro at $25 a month, not $16, and the Analyst tier at $90 was missing from our page entirely. Whether they refill, we no longer claim to know, and an earlier version of this entry did. It said the vendor describes the 15 credits as one-time. The vendor does not use that word; it was our reading of a plan card that omits the per-month wording the paid cards carry, and the billing FAQ lower down the same page calls the free plan a monthly allowance. The review now prints both and resolves neither. We cannot tell from here whether the vendor changed the ladder since July or our July reading was wrong. The pricing subscore moves 3.5 to 3.3, weighing the free tier's ambiguity against a yearly column we had not read until now, which turns out to be published in full and to reconcile exactly; the overall rating is unchanged.
- How it was found.
- The same pre-release check, extended to the AI data analysis cluster.
- What we published.
- An Ultra tier at $500 a month, and a reading of Business at $450 in June against $375 in July as a price cut, with an argument about tier consolidation built on it.
- What is right.
- julius.ai/pricing read on 5 September 2026 carries no $500 tier, and the figure is withdrawn rather than restated. The $450 and $375 are not two dates, they are one plan's two billing states: $450 month-to-month and $375 a month billed annually. The price never moved, so the consolidation argument rested on a billing basis we had misread, though the narrower point survives and is dated on the page: a separate $750 Growth tier present in the June snapshot was absent from both later reads. Every tier now states both bases, and the vendor's advertised 20% yearly discount is exact only on Plus.
- How it was found.
- The same pre-release check.
- What we published.
- From 7 June to 3 September 2026, a paid ladder of $12, $18 and $36 a month, published as flat monthly prices.
- What is right.
- Those are the annual-billing rates. Month to month the plans are $18, $27 and $48, read on humanizemy.ai/pricing on 3 September 2026, a page the vendor marks last reviewed 31 August 2026. A reader who signed up monthly on the strength of the old figures paid 50% more than this page said. The page now gives both bases for every tier. This entry logs a fix made on the page on 3 September that was not written to this log at the time; the same error on the WriteHuman review is logged here twice, and this one should have been.
- How it was found.
- A pre-release check of the AI humanizer cluster, reading the vendor pricing page in a browser.
- What we published.
- From 7 June 2026, that the free tier gave 125 words per run, four runs a day, a maximum of 500 words a day, and one pricing row described the allowance as recurring.
- What is right.
- HumanizeMyAI's pricing page says the opposite in as many words: a free account gets four humanizations of up to 250 words, and "that allowance is a one-time trial rather than a daily refill". A reader planning unpaid ongoing work at 500 words a day was planning on an allowance that does not exist. The page now states the one-time grant, the argument for upgrading is rebuilt around having four runs in total rather than a daily ceiling, the free-tier score moves 3.0 to 2.5, and the vendor's genuinely free AI detector is named instead. Read on humanizemy.ai/pricing, 4 September 2026. The same claim was corrected on the AI humanizer ranking and on the Undetectable AI, Humbot and QuillBot pages, where it had also inverted a comparison in this tool's favour.
- How it was found.
- A pre-release check of the AI humanizer cluster, re-reading every vendor pricing page in a browser before shipping the batch.
- What we published.
- Month-to-month prices of $18, $27 and $48 and annual rates of $12, $18 and $36, read on 19 August 2026.
- What is right.
- WriteHuman has raised prices. Read again on 4 September 2026, month-to-month is $20, $29 and $59, and annual is $12, $19 and $39, charged $144, $228 and $468 a year. Five of the six paid figures moved; only Basic's annual rate held. This one is the vendor's change and not a repeat of the reading error corrected on 19 August, and the page says which is which. Sixteen days separate the two reads, which is why the page now carries the date its prices were last checked. The correction note dated 19 August 2026 in this log has had the clause "on the ladder as it stood then" added to it, because the sentence read as a statement about today's prices; no figure or date in it was altered.
- How it was found.
- The same pre-release check.
- What we published.
- Monthly prices of about $12, $14.99 and $22.99, carried since publication with no billing basis stated.
- What is right.
- Those figures cannot be verified and are withdrawn rather than restated. Read in a full browser on 4 September 2026, humbot.ai/pricing shows no prices at all to a visitor who is not logged in: every plan renders a placeholder where the figure belongs, beneath a banner advertising up to 86% off, so any number quoted from that page would have been a promotional one. The review now describes the tiers by the allowances and input limits the vendor does publish, and says that a price comparison against the rest of the category cannot be made. Why the page behaves that way we do not know and do not guess.
- How it was found.
- The same pre-release check.
- What we published.
- A review, live since 21 July 2026, presenting LOVO AI as an operating product with a 14-day trial, paid plans from $24 a month and a fourth place in our AI voice ranking.
- What is right.
- LOVO, Inc. had filed for Chapter 7 liquidation on 27 May 2026, two months before we published, and every page of lovo.ai has returned an HTTP 402 error on every check since August. The pricing had been read from archived snapshots, so a product in liquidation was written up as a live one. The page now opens with a dated note, is marked discontinued, its calls to action point at a live alternative, LOVO leaves the voice ranking, and every page that compared against it says it is gone.
- How it was found.
- Our research desk recorded the filing on 10 August 2026 while auditing product status for the pricing transparency study; a site review on 2 September noticed the review still treated the product as live.
- What we published.
- The homepage rankings index counted OpenAI Sora, shut down in April 2026, as one of six ranked video tools.
- What is right.
- The Sora review already said the product was discontinued; the flag the index reads had not been set. It is set now, so the index counts it as archived.
- How it was found.
- The same site review.
- What we published.
- The data-availability section on the site still said no DOI had been minted for the release.
- What is right.
- A DOI was minted on 1 September 2026 and the paper's source draft named it the same day; the data-availability sentence in the site copy of the paper had not been updated. The section now names the concept DOI and the v1.0 DOI, and keeps the history of the earlier sentence. Logged as D-087 in the study's deviations log.
- How it was found.
- A site-wide review the next morning, reading the live page against the study's own deviations log.
- What we published.
- The citation file shipped with the public dataset stated 82 dated deviations, and the README stated 78, when the log held 84. The archived v1.2 record on Zenodo was built from that citation file, so a permanent identifier described an audit trail with the wrong count.
- What is right.
- Both files carry the derived count, the archived record was corrected by hand, and both files are now among the documents the release's automated figure check reads. Logged as D-086 in the deviations log.
- How it was found.
- While preparing the archive deposit, by re-deriving the count from the log.
- What we published.
- The rating label described Veo as offering "the only native audio" among the video generators we cover.
- What is right.
- Our own Kling AI and OpenAI Sora reviews document native audio on both. What sets Veo apart is that its audio is not metered separately, and the label now says that.
- How it was found.
- A scan of every review's metadata for uniqueness claims, checked against our own published pages.
- What we published.
- That the $10 Basic plan buys "about 200 fast images", stated wherever the plan's allowance appeared.
- What is right.
- Basic carries 200 fast GPU minutes, not 200 images, and one prompt returns four images. Midjourney publishes no images-per-plan figure; at its documented rate the allowance is worth several hundred images before upscales and re-rolls. The error understated what $10 buys. Every instance now states the minutes and explains the conversion.
- How it was found.
- Research for the Midjourney video guide, reading the allowance against Midjourney's own documentation.
- What we published.
- That the checker runs entirely in your browser. The page title said "In-Browser" and the privacy dimension scored a flat 5.0 on that basis.
- What is right.
- A small function on the vendor's side fetches the target site's robots.txt and returns it; the parsing happens in the browser. The body had already said so, but the title and the score had not moved. The title now says no-log fetch, and privacy scores 4.5, because the no-record assurance rests on the vendor's own statement.
- How it was found.
- Our own review of the page after an earlier partial correction, when the title and score were found still carrying the old claim.
- What we published.
- Every review printed "Pricing verified" followed by the date the page last changed. For 39 of the 77 live reviews the pricing had not been re-read on the vendor's page by that date.
- What is right.
- The stamp now appears only where someone on our side read the figures on the vendor's own pricing page, and states that date. Pages without a recorded check show no stamp in either direction.
- How it was found.
- Comparing the printed stamp against the internal verification notes behind each page.
- What we published.
- A last-updated date of 29 August 2026 on pages that went live on 28 August, so the sitemap advertised a date one day in the future.
- What is right.
- The date is the day a reader could see the change. All six now carry 28 August, and the release tooling refuses a future date.
- How it was found.
- A check of sitemap dates against the day the pages actually shipped.
- What we published.
- From 7 June to 19 August 2026 the page listed WriteHuman's plans at $12, $18 and $36 a month.
- What is right.
- Those are the annual-billing rates. Month to month the plans cost $18, $27 and $48 on the ladder as it stood then. Archived captures from March, June and August show the same ladder, so this was our error and not a vendor change. The page names the old figures rather than replacing them silently.
- How it was found.
- A scheduled refresh that re-read the vendor's pricing page.
- What we published.
- That the release was "published under CC BY 4.0 with a DOI minted at publication". No DOI had been minted.
- What is right.
- The sentence had been carried from the pre-registered protocol, where it was a plan, into the paper, where it read as a fact. It was replaced the same day with a statement of what did identify the release. Logged as D-081; the DOI itself followed on 1 September.
- How it was found.
- A read-through of the paper's provenance claims on publication day.
- What we published.
- A cost table putting Colossyan at $0.18 to $0.34 per rendered minute on a Pro plan, concluding that it undercuts Synthesia above 60 minutes a month. The page also priced Starter at about $19 a month and listed SCORM export as Enterprise-only.
- What is right.
- The per-minute figures came from an allowance that had not been verified; the real arithmetic is about $1.97 a minute on NEO, and Colossyan does not undercut Synthesia per rendered minute at any published tier. Starter is now free, and SCORM export starts on the $59 Professional tier. The conclusion was withdrawn and the old claims are named on the page.
- How it was found.
- A scheduled refresh that re-read Colossyan's pricing and documentation.
- What we published.
- A Team plan at roughly $69 a seat a month, no mention of the Pro or Business tiers, credits that expire monthly with no rollover, and a real-time Interactive Avatar API presented as part of the subscription.
- What is right.
- The Team tier no longer exists under that name, Pro and Business are the tiers that matter, HeyGen's pricing FAQ states that unused credits roll over, and the API is priced separately. One of the four claims was already wrong on the day the page was first published. The rating moved from 4.2 to 4.1, and the avatar ranking's HeyGen-versus-Synthesia call was rewritten on the same day.
- How it was found.
- A scheduled refresh that re-read HeyGen's pricing page and FAQ.
- What we published.
- That AutoShorts.ai launched in September 2025, and that its free plan was a working free tier. The headline rating of 3.8 also never matched the page's own five axis scores, which averaged 3.6.
- What is right.
- Dated public discussion of the product runs back to October 2024, and the founder described reaching $1M ARR that month, which places the launch in the first half of 2024. The free plan is one video with no auto-posting. Both are stated on the page, and the rating now reads 3.6, matching the scorecard.
- How it was found.
- A scheduled refresh that re-checked the product's history against dated public sources.
- What we published.
- The Scale tier at $330 a month, a figure that circulated in earlier coverage including ours.
- What is right.
- Scale is $299 a month, or $249.17 a month billed annually. The page notes the correction beside the tier.
- How it was found.
- A scheduled refresh that re-read the vendor's pricing page.
- What we published.
- That InVideo has no AI avatar, and that dedicated avatar tools win presenter video by default.
- What is right.
- InVideo shipped an AI avatar and talking-head generator in 2026. The page now frames the choice as a trade-off in presenter quality rather than an absence.
- How it was found.
- A scheduled refresh that re-read the vendor's feature documentation.
- What we published.
- Four pages claimed hands-on testing we do not run: the voice hub said we sign up and test before publishing, the faceless-video description opened "We tested", the humanizer hub's structured data claimed ranking by hands-on detector results, and the paid-submission tier promised a hands-on look. Three page titles and fourteen share-card kickers read "Tested & Ranked".
- What is right.
- Our methodology page had promised we would never claim tests we did not run. All four claims were removed, the titles and kickers now read "Compared & Ranked", and the voice hub states plainly that no private lab test was run. The sweep covered hubs, titles and share cards, not review bodies: three reviews (Murf AI, InVideo AI, ElevenLabs) still opened with a hands-on claim and were corrected in their refreshes on 22, 27 and 29 July, and the faceless-video hub's sign-up-and-run line was removed on 31 July.
- How it was found.
- An audit of hub pages, titles and share cards against the promise on /methodology; the review-page instances surfaced one by one in the refreshes that followed.
- What we published.
- A plan ladder of Creator Lite and Creator Plus at $19 and $33 a month, Business at $26 to $52, and Enterprise from $75.
- What is right.
- None of that matched Murf's pricing page: Enterprise is a custom quote rather than a published entry price, and voice cloning is an add-on on top of it. The page carries the corrected ladder and a dated note on the old one.
- How it was found.
- A scheduled refresh that re-read the vendor's pricing page.