Free AI Model Recall Checker
Two measurements in one tool. First, we take the unbranded questions people actually type in your category and send them to more than one free AI model — then count how many of those models name your brand without being told about it. Second, we hand you the same questions with one-click copy and open buttons for ChatGPT, Perplexity, Google AI Mode and Copilot, and keep your own “was I cited?” tally in your browser. The first measures what models remember from training; the second measures the real engines, with you as the operator.
“best analytics tools for startups”
Named by gpt-oss-20b · not named by qwen3.8-27b
“privacy-friendly Google Analytics alternative”
Named by gpt-oss-20b · not named by qwen3.8-27b
“self-hosted product analytics platform”
Named by neither model
Asked: gpt-oss-20b (trained on data up to June 2024); qwen3.8-27b (training cutoff not published). Your brand was never mentioned in the questions, so a mention is the model recalling you.
“best analytics tools for startups”
Your ticks stay in this browser. We never see them, and we never ask those engines anything on your behalf.
Half one: do AI models name you when nobody mentions you?
One AI model writes the questions a person in your category would actually type — unbranded, so your name never appears in them. Those questions then go out to more than one free AI model, and we count the answers that name your brand unprompted. The result reads “N of M models named you, unprompted”, and lists which model named you on which question.
That count is the point. The old version of this tool asked one model to grade itself — to say whether it would mention you — and a model’s opinion of its own knowledge is not a measurement: when language models put a confidence on their own answers they “tend to be overconfident” (arxiv.org/abs/2306.13063, ICLR 2024, read 24 September 2026). Counting real mentions in real answers is repeatable; a self-assessment is not. So the self-graded score is not demoted, it is gone — along with the model’s own “recognised / confidence / sentiment” labels and its guesses about which prompts it would name you on. When the count cannot be taken, the result says that, instead of falling back on a number a model made up about itself.
What this half is not:
- It is not a live search. Nothing is retrieved from the web during the check. It is what those models absorbed during training.
- It is not ChatGPT, Claude, Perplexity or Copilot. We send nothing to them. A strong result here does not promise a citation there.
- It is a small sample. A few questions, a few models, one run each. Vendor tests reported by Search Engine Journal found identical prompts returning the same brand list less than 1% of the time (searchenginejournal.com, 30 January 2026, reporting SparkToro and Gumshoe, read 24 September 2026). Read a zero as “thin memory”, not as a verdict.
- Training data has a cutoff. Each model’s cutoff is shown with the result. Anything that happened to your brand after it cannot be there.
Half two: check it yourself in the real engine
The same unbranded questions appear again with a copy button and open buttons for ChatGPT, Perplexity, Google AI Mode and Copilot. You run them, you tick the engines that named or linked you, and the panel keeps a running tally — your own spot-check: 3 of 12. Run each question twice: the same prompt rarely returns the same brands twice, so one run proves very little.
Why we ask you to do it. We cannot query those engines for you. Doing it from our servers would mean scraping results, buying a paid quota, or working around someone’s terms — and this site is free, has no paid AI budget, and is not going to pretend a simulation is the real thing. Handing you the questions is the honest version: it is the actual engine, answering you, in your country, with your account.
Your ticks are yours. They are stored in your own browser and never sent to us — no account, no upload, and nothing about you kept on our side. Clearing your browser data clears the tally, and it will not follow you to another device.
If an engine opens with an empty box, paste. None of these engines documents a link format that pre-fills a question, and they change them without notice — which is why every question has a Copy button next to the open buttons. The copy is the part we can promise.
Which models answer, and why only a handful
Today the questions go to two free models: OpenAI’s gpt-oss-20b — open weights under a “Permissive Apache 2.0 license”, with a stated “Jun 01, 2024 knowledge cutoff” (developers.openai.com, read 24 September 2026) — and qwen3.8-27b, for which we have no published cutoff we can cite, so your result says “training cutoff not published” rather than inventing one. Trust the model list in your result over this paragraph: the roster is configurable, and a busy provider can route a request elsewhere, so the result names whichever models actually answered and the cutoff of each.
The limit is quota, not ambition. The free tier we run on allows 30 requests and 8,000 tokens per minute, and 200,000 tokens per day, per model, across the whole organization (console.groq.com/docs/rate-limits, read 24 September 2026) — and six other AI features on this site draw from the same pool. So the check is built to fit: three questions to two models is six small calls, not the eighteen a six-question, three-model fan-out would need; the probe models are deliberately ones the rest of the site does not use, so a brand check cannot starve the audit assistant; a brand is probed once a day and repeat lookups come from that day’s cache; and a site-wide daily ceiling stops one busy day spending the quota everything else needs. That ceiling is a single date-keyed number — no IP, no session, nothing per visitor.
If fewer models answered than we set out to ask, the result says so and counts only what ran: two of two, or one of one with the shortfall spelled out. When the daily ceiling is spent, or the free models are at capacity, you get no number at all and the questions instead — because a partial fan-out presented as the full check is exactly the dishonesty this rebuild removed.
One case we refuse to score. If your brand name is an ordinary word — Monday, Apple, Stripe — counting how often a model says it measures nothing, because the model is usually talking about something else. Give us your website and we count mentions of your domain instead, which is unambiguous; without one, the result says it cannot measure this rather than handing you a number nobody should believe.
Where the real numbers live
Neither half of this tool counts citations across the whole web, and no free API exposes that. Your own first-party reports do: Search Console → Generative AI (impressions in Google’s AI Overviews and AI Mode), Bing Webmaster Tools → AI Performance (citations in Copilot and Bing AI answers, which our AI Citation Report reads in your browser), and GA4’s “AI Assistant” channel (visits that arrived from an assistant). Those are real numbers about your site; this tool is not a substitute for them.
What tends to help
- Let AI crawlers fetch your site — check with the AI Bot Access Checker.
- Put the substance in your server-rendered HTML, not behind JavaScript — see Bot-Eye View.
- Name your organization and products the same way everywhere, and keep the pages current.
- Earn genuine coverage on sites people already read. Google warns explicitly against arranging inauthentic mentions, so buying them is a risk, not a shortcut.
- Run a full SEO and AI-readiness audit for what we can actually check on your site.
Frequently asked questions
Does this check what ChatGPT, Claude or Perplexity say about my brand?
Not from our side, and we say so rather than implying otherwise. We ask free AI models — named in your result, with their training cutoffs — the unbranded questions, and count how many name your brand unprompted. That is training memory. Querying ChatGPT, Perplexity, Google AI Mode or Copilot from our servers would mean scraping, a paid quota or a terms problem, and we do none of those. That is exactly why the second panel exists: it gives you the same questions with copy and open buttons so you can run them in those engines yourself, and tick what you see. Run each question twice, because their answers change between runs.
What does “1 of 2 models named you” mean?
Two free AI models were asked the same unbranded questions — questions that never contain your brand name, your website or a product only you sell — and one of them named your brand in at least one answer without being prompted. The denominator is always the models that actually answered, never the models we meant to ask, and the result lists which model named you on which question along with each model's training cutoff. It measures what those models remember from training. It is not a prediction about any AI product, and it says nothing about whether a real engine would cite your page today.
Why only three questions and two models?
Free quota, shared across the whole site. The free tier we run on allows 30 requests and 8,000 tokens a minute and 200,000 tokens a day per model for the whole organization (console.groq.com/docs/rate-limits, read 24 September 2026), and six other AI features here draw on the same pool. Three questions to two models is six small calls; six questions across three models would be eighteen, and would spend a minute's whole budget on one brand check. We also probe models the rest of the site does not use, cache a brand for a day, and keep a site-wide daily ceiling that is one date-keyed number with nothing per visitor in it. The owner's rule is that this site stays free with no paid AI models, so the honest response to a limit is a smaller check that reports exactly what it did.
Where are my “was I cited?” ticks stored?
In your own browser, and nowhere else. There is no account, nothing is uploaded, and we store nothing about you — not your ticks, not your IP. The tally lives with this browser on this device, so clearing your site data clears it and it will not appear on your phone. If you want to keep a record, copy the questions and your notes somewhere of your own.
My brand name is an ordinary word. Can you still measure it?
Only if you give us your website, and then we count something different. A model that says “monday”, “apple” or “stripe” is usually not talking about your business, so counting mentions of the word would measure the English language rather than your brand. With a website we count mentions of your domain instead, which is unambiguous, and the result says that is what was counted. Without one we report that we could not measure it — which is not the same as saying no model named you, and the result never shows those two as the same thing.
No model named my brand. What should I do?
It usually means there is not much authoritative, public information for models to have learned from — and remember every model has a training cutoff, so anything recent about you cannot be in there yet. What plausibly helps: making sure AI crawlers can fetch your site, putting the substance of your pages in server-rendered HTML, keeping your naming consistent, and earning genuine coverage on sites people already read. These shift slowly, and only affect models trained after the change, so don't expect a different answer next week. Meanwhile the spot-check panel is the more useful half: the real engines can retrieve pages live, so they may cite you even when a model's memory has nothing.
Does improving this count help my Google rankings?
Not directly, and neither number is a ranking factor anywhere. The underlying work — crawlable pages, substance in the HTML, consistent naming, genuine coverage — supports classic search too, which is why we recommend it. But the count measures what a handful of models remember, and the spot-check measures what you saw in an engine on the day you looked.
Related free tools
Instant SEO Snapshot
One-click SEO & AI-readiness grade for any page — HTTPS, indexability, title, meta, headings, content & more. Embeddable on your own site.
AI Bot Access Checker
What does your robots.txt say to OAI-SearchBot, PerplexityBot, Claude-SearchBot, GPTBot and the rest? Read with the real rules.
NLP Content Analyzer
Topics, entities, search intent & content gaps an NLP model sees in your page.
AI Retrieval Checker
AI-retrieval readiness for one page — an AI model’s read of how findable and quotable it is, with fixes.
AI Crawler Log Analyzer
Upload your access log to see which AI bots (GPTBot, ClaudeBot, PerplexityBot) actually crawled you — with IP verification where the operator publishes ranges. Runs in your browser.
AI Citation Report (Bing Copilot)
Upload your Bing Webmaster AI Performance export to see which of your pages Microsoft Copilot and Bing’s AI answers cite, and how that trends. Covers Copilot and Bing only — Google’s AI numbers are in Search Console’s own Generative AI report. Runs in your browser.
Want the full picture?
This tool checks one thing. Run a complete, free SEO audit across 28 modules.
Run a free SEO audit