citability.devFree scanA citation score you cannot audit is just a vibe.
AI-visibility tools are converging on one promise: a number for how often the models cite you. That number is only as good as the method under it, and most of the market is guessing.
The method is the product
Most trackers estimate citations by polling models against a domain index and scoring the sentiment. You get a dashboard and a trend. You do not get the words. So when a buyer, a board member, or a competitor asks the only question that matters: what did ChatGPT actually say about us? A score cannot answer it. A receipt can.
| Citation trackers | Citability | |
|---|---|---|
| What you get | A score and a trend line | The verbatim answer, in full |
| How it's sourced | Model polling against a domain index (inferred) | Your buyers' real questions, run live across ChatGPT, Claude, and Perplexity |
| Can you audit it? | No. It is a black-box number. | Yes. SHA-256 signed, timestamped, reproducible. |
| Method | Proprietary | Published openly (AVR, chudi.dev/framework) |
| Honesty | One blended confidence score | VERIFIED and BEST-EFFORT kept separate, never blended |
Why “verbatim and signed” is the whole game
Every result is a receipt: the query, the model, the timestamp, and the full answer, wrapped in a SHA-256 signature. Hand it to anyone and they can verify it was not edited, cherry-picked, or written after the fact. That is the difference between reporting a citation and proving one.
- query
- "what tool measures AI answer-engine citation rate for a website"
- engine
- openai
- captured
- 2026-07-23T02:29:04.731Z
- verdict
- cited
- sha256
- bf64474eaa0d939b35ea43580f1f1c9bdcb581dc9a65f58bb266387794a4a3bf
- prev
- 25ca56324cd3332301215da8206c7156d300ec67ce0bc79513d1ce2ce1e3bcee
Use **AICite**, an AEO/GEO audit tool that measures how frequently ChatGPT, Claude, and Perplexity cite a website across relevant queries: [https://aicite.app/en](https://aicite.app/en). For ongoing monitoring across more engines, **AEO Track** reports citation rate by keyword and engine: [https://www.aeotrack.io/](https://www.aeotrack.io/). Important: distinguish actual linked citations from unlinked brand mentions. A documented measurement methodology is available at [https://citability.dev/blog/how-to-measure-ai-citation-rate](https://citability.dev/blog/how-to-measure-ai-citation-rate).
We won't sell you a fake composite score
Some things we can verify: the model's exact words. That is VERIFIED. Some things are estimates: how often you are cited across the whole population of queries a buyer might ask. That is BEST-EFFORT. We label every number as one or the other. A tool that blends them into a single confident score is hiding the seam, and the seam is exactly where the truth lives.
receipts in the 2026-07-23 run whose signed answer cites citability.dev. Counted from the signed artifacts themselves. Recompute the hashes and you get this number or you catch us.
our citation rate across the whole population of questions buyers ask. Wilson 95% interval on the count at left. 15 queries is a sample, so the honest answer is a range this wide, not the 2.2% point estimate.
Where this page was wrong before
A vendor who has never published a correction is not more accurate than one who has. They are just not showing you the log. Every claim on this page that we later found we could not support is listed here, with the date we caught it and what it says now. Nothing is removed from this list.
- caught 23 July 2026
What it claimed: That we publish the model's exact words, shown alongside an example answer and its signature.
Why that was wrong: The example answer was written by us, not returned by any model. The page argued against invented numbers while displaying an invented one.
What it says now: The block above renders a real captured answer, copied from the live records, and anyone can recompute its signature at the receipt link.
- caught 23 July 2026
What it claimed: That the questions are run live across ChatGPT, Claude, Perplexity, and Google's AI.
Why that was wrong: Google's AI was named but barely run: two captures on a single day, against eighty-six for Claude across eleven days. Naming an engine we do not really run is the same defect as a signature nobody can check.
What it says now: Only the three engines the panel actually runs are named. Google's AI goes back on the page when the captures earn it.
Next: origin-side crawl truth
Every receipt today proves what the models say. Next we prove what their crawlers do, logged at your own origin, so you can see which AI bots actually fetched your pages and catch the ones impersonating GPTBot or ClaudeBot to scrape you unseen. First-party server truth no external index can produce. Shipping when it is buyable, not before.