Skip to content
Growzo · AI visibility

Four AI engines. One real citation signal.

Growzo probes four AI engines — Perplexity, ChatGPT, Claude and Gemini — and only one of them can prove a citation. Perplexity runs on sonar-small-online with live web grounding, and Growzo parses the citations array its response actually returns. ChatGPT, Claude and Gemini answer from training knowledge, so their source lists are synthesized by extracting hostnames from the answer text — an inference, labelled as one, and never counted as a verified citation.
Diagram: a tracked prompt passes a budget gate and is dispatched round-robin to four AI engines. Perplexity is live web-grounded and its citations array is parsed into real cited URLs; ChatGPT, Claude and Gemini answer from training knowledge, so their sources are synthesized by extracting hostnames from the prose. With no LLM key, a labelled Brave top-five serp-proxy is used instead and excluded from the citation count.
One prompt, four engines, two very different kinds of evidence — kept apart on purpose.
01 · What each engine proves

Four engines, and the line between citation and inference.

Every AI-visibility tool shows you a list of sources next to an answer. Almost none of them tell you where that list came from. It matters enormously: an engine with live web search hands back the URLs it actually consulted, while an engine answering from training weights hands back prose that merely mentions some domains. Growzo keeps the two apart in the data model, not just in a footnote, so a synthesized source can never quietly inflate the number of engines citing you.

EngineModelGroundingWhere the sources come fromVerified citation
Perplexitysonar-small-onlineLive web searchParsed from the response's citations arrayYes
ChatGPTgpt-4o-miniTraining knowledgeSynthesized from hostnames in the proseInference
ClaudeHaikuTraining knowledgeSynthesized from hostnames in the proseInference
Geminigemini-1.5-flashTraining knowledgeSynthesized from hostnames in the proseInference
Brave serp-proxyNo model — fallbackBrave top-5 resultsSearch positions, labelled as a proxyNo
Perplexity is the only row that can answer “did an AI engine cite this page?” with evidence. The three training-knowledge rows still answer a genuinely useful question — whether a model, unprompted, associates your brand with a buying question — which is a real GEO signal, just a different one. The last row is the no-key fallback, described in full further down. Growzo does not probe Copilot, Grok, You.com or Phind, and it does not detect Google AI Overviews.
02 · What a probe records

Every probe stores six things, on every engine.

A mention is not a boolean. Growzo keeps the full answer text alongside whether your brand was mentioned at all, where you sat in the ordered source list, that list itself, a sentiment score for how you were described, and the word associations the engine reached for. That last one is the quietly useful field: knowing you are cited is one thing, knowing you are cited as “cheap” rather than “compliant” is what changes the next page you write.

Probe run — one tracked prompt
Budget checked · ₹0.20 probeallowed
Perplexity · sonar-small-onlinecited, position 2
ChatGPT · gpt-4o-minimentioned, synthesized
Claude · Haikunot mentioned
Gemini · gemini-1.5-flashmentioned, synthesized
Round-robin dispatch — one slow engine cannot starve the other three.

Budget-gated before it runs

Every probe, every SERP call and every model call is checked against a hard money cap before it happens, and metered after. It is a stop, not a dashboard: at 80% of the cap Growzo alerts, and at 100% it refuses the call rather than continuing and billing you for the discovery. Daily probe caps are set per tier, and the published unit economics are computed from the same constants the ledger bills against.

Round-robin, not sequential

Engines are dispatched in rotation rather than one queue at a time, so a slow or rate-limited provider delays its own results instead of holding up the run. A day where one engine is degraded still produces three engines of usable data.

03 · One number

An AI Brand Score you can recompute by hand.

Growzo reduces the four engines to one 0–100 figure over a rolling 14-day window: the mean of your mention frequency multiplied by the mean of your position-inverse, times 100. Being mentioned often but last scores badly; being cited first but rarely scores badly too. Both halves have to be true, which is exactly how an AI answer behaves — the third source in a five-source answer is read far less than the first.

AI Brand Score — 14-day window
Perplexity72 · cited
ChatGPT54 · inferred
Gemini41 · inferred
Claude28 · inferred
Per-engine breakdown, so a single grounded engine never hides behind three inferred ones.

The delta uses the same function

The previous-period comparison is not a second calculation. It calls the identical scoring function with the window shifted back, so the two numbers can never come from different formulas — the failure that makes most “up 12%” badges meaningless. If the score moved, the inputs moved.

Share of voice, per host

Over a 30-day window Growzo aggregates which hosts appear as sources for your tracked prompts and what share each one holds. It is the map of who currently owns the answer you want, and it is usually a shorter list than the search results for the same question.

4
Engines probed
Perplexity, ChatGPT, Claude and Gemini. Not eight.
1
With real citations
Perplexity returns a citations array Growzo parses.
14 days
Brand Score window
Rolling, with the delta computed by the same function.
30 days
Share-of-voice window
Per-host source share across your tracked prompts.
04 · From signal to work

A citation ledger, and five kinds of recommendation.

Every mention lands in a per-mention ledger — which engine, which prompt, which position, which sources sat above you, and the answer text as it was returned. That ledger is what makes a recommendation arguable rather than mysterious: you can open the row that produced it. From it Growzo emits five kinds of GEO recommendation.

create_comparison

Publish the comparison page an answer would need in order to name you next to the rival it currently names.

earn_citation

Go after a citation from a host that keeps appearing as a source on your tracked prompts.

add_schema

Add the structured data that anchors the passage an engine would otherwise have to guess at.

improve_position

Move up the source list. Position is half the score, so being cited last is worth far less than being cited first.

expand_coverage

Cover a tracked prompt your published corpus does not yet answer at all.

None of the five place a link, buy a mention or contact anyone on your behalf. They describe work a person does, which is the same posture Growzo takes in link building and in indexing — it finds opportunities and refuses to fake the outcome. The Growzo overview covers the rest of the workflow those recommendations feed into.

05 · Honest degradation

No LLM key? You get a proxy, clearly labelled as one.

Growzo is bring-your-own-key, and plenty of people evaluate it before adding one. Rather than showing an empty page, it builds a serp-proxy row from Brave’s top five results for the tracked prompt: position one scores 1.0 and position N scores 1/N. That is a reasonable stand-in for visibility and a terrible stand-in for a citation, so it is treated as exactly the first and never the second.

Labelled everywhere it appears

The row carries its provenance through the UI, so no chart can present a Brave ranking as though a model had quoted you. A visibility proxy that forgets it is a proxy is how a dashboard ends up lying to the person paying for it.

Excluded from the citation count

Proxy rows are left out of “LLMs citing you” entirely. The count means engines that returned a parseable citation, which today means Perplexity — and if that number is zero, Growzo shows zero.

Rank tracking is Brave only

One search engine, not seven. Google, Bing, Yandex, Baidu, DuckDuckGo and Naver have no implementation in Growzo, and AI Overview presence is a hard-coded placeholder rather than a measurement. Whether an engine can reach your pages at all is a separate question, answered on the AI crawler page.

Open source, self-hosted, pre-1.0

You can read the prober and count the engines yourself. That is the point of stating four instead of eight: on a product a buyer can inspect, an overclaim has a shelf life of about one afternoon. Questions welcome at contact, or see the rest of the product family.

06 · Questions

AI citation questions, answered plainly.

How many AI engines does Growzo actually probe?

Four: Perplexity on sonar-small-online, ChatGPT on gpt-4o-mini, Claude on Haiku and Gemini on gemini-1.5-flash. Only Perplexity is live web-grounded, and only Perplexity returns a citations array that Growzo parses into real cited URLs. The other three answer from training knowledge, so their source lists are synthesized by extracting hostnames from the answer text — an inference, labelled as one, never presented as a citation.

Does Growzo track Copilot, Grok, You.com or Phind?

No. Those names appear in an internal enum with no prober behind them, so Growzo does not track them and this page will say so until it does. Any tool claiming eight-engine citation tracking is worth asking the same question of: which engine returns citations, and which one is being inferred from prose?

Does Growzo detect Google AI Overviews?

No. AI Overview presence is hard-coded false in the rank worker with a comment marking it a deliberate placeholder, so Growzo reports nothing about AI Overviews rather than reporting a zero that would read as absence. Rank tracking itself runs on Brave only — one search engine, not seven.

How is the AI Brand Score calculated?

It is a single 0-100 number over a rolling 14-day window: the mean of your mention frequency multiplied by the mean of your position-inverse, times 100, with a per-engine breakdown underneath. The previous-period delta is computed by the same function with a shifted window, so the current figure and the change can never come from two different formulas — the most common way a trend line lies.

What happens if I have no LLM API key?

Growzo falls back to a labelled serp-proxy row built from Brave's top five results: position one scores 1.0 and position N scores 1/N. It is marked as a proxy everywhere it appears and is excluded from the count of LLMs citing you, because a search ranking is not evidence that a model cited you. Growzo is bring-your-own-key, so the real probes start the moment you add one.

Is Growzo open source, and can I self-host it?

Yes. Growzo is open source, self-hostable and pre-1.0, and it runs on your own API keys. That is also why the numbers on this page are stated precisely: anyone can read the prober, find four engines, and check that the fifth column of the table below matches the code.

Track what an AI answer can actually prove.

Growzo runs on its own application and its own sign-in, on your own API keys. Add one key and the probes start; add none and you get the labelled proxy instead of an invented number.

Start freeTalk to us