Skip to content
Growzo · ChatGPT, Claude, Gemini

Growzo will not tell you ChatGPT cited you. Here is why.

Growzo probes ChatGPT on gpt-4o-mini, Claude on Haiku and Gemini on gemini-1.5-flash — three training-knowledge probes that never retrieve anything, so none of them returns a citations array to parse. Growzo synthesizes a source list by extracting hostnames from the answer prose, labels it an inference, and excludes it from the count of LLMs citing you. What the probe genuinely measures is different and still valuable: whether a model, unprompted, associates your brand with a buying question, where you fall in the order it names names, how it describes you, and which words it reaches for.
The same tracked prompt goes to ChatGPT on gpt-4o-mini, Claude on Haiku and Gemini on gemini-1.5-flash, which answer from training knowledge with no live retrieval and return no citations array; Growzo synthesizes a source list by extracting hostnames from the answer prose, labels it an inference and excludes it from the citation count, while still recording the full answer, the mention flag, the position, a sentiment score and the word associations.
No retrieval means no source list to return. What the probe still records is on the left.
01 · Weights, not retrieval

A chat model has nothing to cite.

When a model answers from training knowledge it is writing from what it learned long before your question existed. It opened no page, followed no link and consulted no index, so there is no list of sources for it to hand back. Anything presented as a source list for such an answer has to be reconstructed after the fact — usually by finding the domains the sentence happens to name. That reconstruction is a reasonable thing to do and a completely unreasonable thing to call a citation, and the gap between those two is invisible on a dashboard that renders both as a tidy row with a favicon.

EngineModelWhat it readsWhere its sources come fromA citation?
ChatGPTgpt-4o-miniTraining knowledgeHostnames extracted from the answer proseInference
ClaudeHaikuTraining knowledgeHostnames extracted from the answer proseInference
Geminigemini-1.5-flashTraining knowledgeHostnames extracted from the answer proseInference
Perplexitysonar-small-onlineLive web searchParsed from the response's citations arrayYes
The last row is the exception that makes the first three legible. Perplexity retrieves before it answers, so Growzo can parse the URLs it consulted; that path is documented separately. The other three are grouped here not because their answers are worthless but because they are worth something different, and flattening the two kinds into one count is how an AI-visibility number quietly stops meaning anything.
02 · The signal that is real

Whether a model already knows who you are.

A training-knowledge probe answers a question no crawl log and no rank tracker can: when a model reaches into what it learned and is asked which tools solve this problem, does your name come out? That is brand knowledge rather than retrieval, and it moves on a completely different timescale — slowly, and mostly in response to how widely and how consistently you are described elsewhere. Growzo records six fields on every probe, and the two that never reach the score are often the two worth reading first.

Training-knowledge probe — one promptmentioned · position 3
Full answer textstored verbatim
Brand mentionedyes, unprompted
Position in the order3 of 6 named
Sourcessynthesized — inference
Sentimentneutral-positive
Word associationsaffordable · self-hosted
Not a citationIllustrative probe. The label on the source row is not illustrative.

Word associations are the quietly useful field

Knowing a model names you is one thing. Knowing it names you as “cheap” rather than “compliant”, or as “for freelancers” when you sell to enterprises, is what changes the next page you write. It is the closest thing in the product to reading your own positioning back from the outside, and it costs one probe.

Sentiment is not the same as presence

Being named in a sentence that goes on to describe you as the limited option is a worse outcome than not being named at all, and a boolean mention flag cannot see the difference. Growzo scores how you were described from the same stored answer text, so the two facts stay attached to each other.

Mentions and positions from these three engines do feed the AI Brand Score — the mean of your mention frequency multiplied by the mean of your position-inverse, times 100, over a rolling 14-day window — but the synthesized source list is never counted as a verified citation. The full derivation is published, including the per-engine breakdown that keeps one grounded engine from hiding behind three inferred ones.

04 · The line nobody else draws

A synthesized source looks exactly like a real one.

Put the two side by side on a screen and they are indistinguishable: a domain, a favicon, a position number. That is precisely why the distinction has to live in the data model rather than in a footnote a product manager can quietly delete. Growzo stores what kind of evidence each source row is, carries that provenance through the UI, and keeps synthesized rows out of the count of LLMs citing you — so no chart can present a hostname found in a sentence as though a model had quoted you.

LLMs citing you — what the count includes1 of 4
Perplexity — parsed citations[]counted
ChatGPT — synthesized sourcesexcluded
Claude — synthesized sourcesexcluded
Gemini — synthesized sourcesexcluded
Brave serp-proxy — no LLM keyexcluded
The mentions and positions from all four engines still feed the AI Brand Score. Only the citation count is this strict.

Strict on one number, generous on the rest

This is the compromise worth understanding, because it is the opposite of the usual one. Growzo will happily use a training-knowledge probe to tell you a model named you third and called you affordable — that is real information and it is reported in full. What it refuses to do is let that probe increment a number whose name promises evidence. One metric is held to the strictest possible standard precisely so the other fifteen can be useful without anybody having to audit them, and so a reader who quotes “one LLM cites us” in a board deck is not quoting an inference.

Excluded from the citation count

“LLMs citing you” means engines that returned a parseable citation. Today that is Perplexity, and only Perplexity. If the number is one, Growzo shows one. If it is zero, it shows zero rather than borrowing three inferred rows to round it up.

The same rule for the no-key proxy

Without an LLM key Growzo builds a labelled serp-proxy row from Brave’s top five results — position one scores 1.0, position N scores 1/N — and excludes it from the citation count for exactly the same reason. A search ranking is not evidence that a model named you.

Four names with no prober at all

Copilot, Grok, You.com and Phind are enum values with nothing behind them. They are not probed, not scored and not counted. Any tool selling eight-engine coverage is worth one question: which of the eight returns a citation?

No AI Overview detection

Google AI Overview presence is hard-coded false in the rank worker as a deliberate placeholder, so Growzo reports nothing about it rather than a zero a reader would take for absence. Rank tracking is Brave only.

Round-robin, budget-gated

The engines are visited in rotation so a slow provider delays only its own results, and every probe is checked against a hard money cap before it happens and metered in paise after. The caps are published with their arithmetic shown.

Readable, because it is open

Growzo is open source, self-hosted, bring-your-own-key and pre-1.0. Every number on this page can be checked against the prober by anyone who runs it, which is the real reason the wording is careful. The overview page carries the full list of what the product does not do.

Nothing here claims or implies a guaranteed mention, citation, ranking or lead — AI systems are proprietary, personalised and query-dependent, and their answers differ between sessions. What Growzo offers is a measurement with its evidence attached and an honest label on the half that is inference. The four-engine matrix puts every row in one table, and contact is open for the awkward questions.

05 · Questions

ChatGPT brand-mention questions, answered plainly.

Does ChatGPT cite my site, and can Growzo tell me?

Growzo cannot tell you that, and it will not pretend to. It probes ChatGPT on gpt-4o-mini, which answers from training knowledge rather than searching the web, so the response contains no citations array to parse. Any source list shown for that probe is synthesized by extracting hostnames from the answer prose — an inference about what the model knows, labelled as one, and excluded from the count of LLMs citing you.

Then what does a ChatGPT brand probe actually measure?

Whether the model names your brand unprompted when asked a buying question, where you fall in the order it names names, how it describes you, and which words it associates with you. Growzo stores the full answer text alongside the mention flag, the position, the source list, a sentiment score and the word associations. That is a genuine generative-engine signal about brand knowledge — it is simply not a citation.

Why can a training-knowledge probe not return citations?

Because the model never retrieved anything. It writes from weights learned before the question was asked, so there is no list of consulted URLs for it to return. Perplexity is different in kind rather than in degree: Growzo probes it with sonar-small-online, a live web-search model whose response carries a citations array that can be parsed into the URLs it actually used.

Is ChatGPT Search the same thing as ChatGPT?

No, and the difference is the whole reason this page exists. ChatGPT Search cites from its own index, crawled by OAI-SearchBot. The probe Growzo runs talks to a chat model over an API and touches no index at all. Growzo tracks the crawler policy that governs inclusion — OAI-SearchBot is one of the thirteen agents in its registry — but it does not measure ChatGPT Search citations, and no page here will imply that it does.

Does blocking GPTBot stop ChatGPT from citing me?

No, and this is the most expensive misunderstanding in the category. GPTBot is OpenAI's training crawler, OAI-SearchBot crawls the index ChatGPT Search cites from, and ChatGPT-User fetches a page live when a prompt causes a browse. Three tokens, three consequences: blocking GPTBot leaves ChatGPT Search citations untouched, while blocking OAI-SearchBot removes you from them entirely.

Does the inferred source list count toward my AI Brand Score?

The mention and the position do; the synthesized source list is never counted as a verified citation. The score is the mean of your mention frequency multiplied by the mean of your position-inverse over a rolling 14-day window, with a per-engine breakdown underneath, so a grounded engine never hides behind three inferred ones. The count of LLMs citing you means engines that returned a parseable citation, which today is Perplexity alone.

Know which half of your AI visibility is evidence.

Growzo runs on its own application and its own sign-in, on your own API keys. It reports the mention, the position, the sentiment and the words — and tells you plainly when the source list was inferred.

Start freeTalk to us