How AI Visibility Platforms Track Competitor Citations

A practitioner's guide to how tools actually pull competitor brand citations out of ChatGPT, Claude, Perplexity, and Gemini, and where the measurement gets messy.

Updated on: 2026-08-12

The first time a client asked me to prove their competitor was "winning ChatGPT," I ran the same prompt three times and got three different answers. Different brands. Different order. One run didn't cite anyone at all. That is the whole problem in one sentence: an AI answer is not a search results page, and treating it like one produces confident nonsense.

AI visibility platforms track competitor citations by repeatedly sending a fixed set of buyer-style prompts to each engine, capturing the raw answer and its sources, then extracting which brands got named, which got recommended, and which domains got cited. The hard part is not sending prompts. It is separating a brand being mentioned from a brand's website being used as evidence, and doing it consistently enough that a month-over-month chart means something.

Most bad reporting collapses everything into one "visibility score." That hides the parts a marketer can act on. There are at least three separate measurement events, and a competitor can win one while losing another.

Peec's documentation makes this split explicit, separating brand visibility (you're mentioned) from source visibility (your domain is used). A company can be cited without being named, or named without its site being touched at all.

Why this matters for competitor tracking: a rival might dominate recommendation lists purely because of third-party reviews and Reddit threads, while barely getting cited from its own site. That tells you where to compete. A blended score erases it.

Everything starts with prompts that sound like buyers, not keywords. "Best AI visibility platform for an agency" beats "AI visibility tool" because it triggers the recommendation behavior you're trying to measure.

A workable prompt set covers category discovery, comparisons, alternatives to a named competitor, audience-specific questions, and problem-framed queries like "how do I find which sites ChatGPT cites instead of mine." One published protocol uses roughly 20 to 25 buyer-intent questions per category and freezes them at baseline. That freezing is the part people skip. If you quietly swap in easier prompts next month, visibility "improves" for reasons that have nothing to do with the work.

For a fair comparison, you send the same wording to ChatGPT, Claude, Perplexity, and Gemini, and you log everything around it: exact prompt text, prompt ID, engine, model version if exposed, timestamp, country, language, session state, and search mode. These four products don't retrieve the same way, so context you forgot to record becomes noise you can't explain later.

This is the line I'd draw between a defensible platform and a black box. Store the actual answer text, the brands named, their positions, the recommendation framing, every visible source URL, the cited page titles, and the parser version. When a client challenges why a competitor got credit, you open the receipt. seoforgpt is built around exactly this: it shows the exact answers, the rankings, the competitors sitting behind each visibility score, and the pages used as evidence, so you can open a prompt where a rival took the shortlist spot you wanted. If a tool only hands you a percentage, you can't audit a single false positive.

Brand extraction combines exact-name matching, alias and entity matching (parent company, old name, domain), domain matching, and context classification that decides whether the brand was recommended, criticized, or just used as an example. String matching alone will confuse similarly named companies and generic category terms, so the good systems keep a human or model check for the ambiguous cases.

Then URLs get normalized. The same source shows up as `example.com/page`, `www.example.com/page/`, and a version dragging tracking parameters. A citation system should collapse protocol, www, trailing slashes, params, and redirects into one canonical page while preserving the original as evidence. Report both page-level citation share (which exact URLs win) and domain-level share (which brands win overall). Profound's citation API is a concrete example of this, breaking citations down by domain, page, model, topic, region, and prompt.

The four engines don't hand over evidence the same way, and this changes what you can actually measure.

ChatGPT shows inline citations or a separate Sources panel. Behind the scenes it may rewrite your question into several search queries, run follow-up searches, and factor in location. So the visible prompt is not necessarily the retrieval query. Capture the answer and the source panel together. Separately, OpenAI documents `OAI-SearchBot`; sites that block it won't appear in ChatGPT Search answers, though that's a crawl-access issue, not a citation guarantee.

Claude decides per prompt whether to search, and its documented citation object is unusually clean: source URL, source title, and up to 150 characters of the actual cited text. That excerpt lets you distinguish "the domain appeared somewhere" from "this specific passage supported the claim." The catch is that Claude often answers stable questions without searching at all. When that happens, label the run "no web search" rather than recording it as the brand being invisible.

Perplexity can be measured at two layers. The Agent API returns web-grounded answers with built-in citations; the Search API returns raw ranked results with titles and URLs. Don't conflate them. A page can sit in raw search results without ever making the final answer. Treat Perplexity as citation-rich, but don't claim every returned source shaped the response unless the API proves it.

Gemini with Google Search grounding is the most structured for source-level work. When grounding is on, it generates search queries, retrieves results, synthesizes, and returns citations with start and end indexes mapping answer spans to specific URLs. That means you can see both what it searched and which sentence each source backed. One caveat worth stating in any report: specify whether you tested Gemini Apps, the Gemini API, or Vertex AI, because they aren't automatically identical. Google's `Google-Extended` token controls certain Gemini uses and does not affect ordinary Search ranking.

If you're focused on one engine, this deeper walkthrough on tracking brand mentions in Gemini goes further than I can here.

One number is convenient and usually misleading. I'd rather report several independent metrics.

The share-of-voice denominator is where people quietly fudge things. Does it mean all brands mentioned? All recommended? All domains cited? Weighted by position? Each choice produces a different chart. A platform should disclose which it uses.

Position matters too. A first-slot recommendation is worth more than a name buried in a paragraph, so tools apply weights (first recommendation 1.0, second 0.8, unranked mention 0.25, source-only citation tracked separately). Those weights are editorial choices, not physics. They belong on the surface of a report, not hidden inside a "visibility score."

Blending the four engines is where actionable detail goes to die. They cite different pages and recommend different brands.

The real reporting unit is `prompt × engine × run × brand × source`. Roll up to category or time-period scores only after you've kept that granular data. seoforgpt leans into model-level reporting for this reason, showing where a brand is winning, absent, or just needs different work per assistant.

The same prompt on the same engine returns different answers minutes apart. Research summarized by Search Engine Land found AI recommendation lists were repeated identically less than 1% of the time, and matched in the same order closer to 1 in 1,000. So a single answer is a coin flip you mistook for a ranking.

A serious tracker runs each important prompt multiple times, reports the number of runs, and calculates appearance frequency instead of a yes/no.

That reads as "you're stable in Perplexity, transient in Claude, and Competitor A owns discovery." A single "65% visibility score" tells you nothing you can fix.

Good competitor tracking answers more than "who showed up." For every prompt, record the recommendation order, whether the answer used a numbered list, positive or negative framing, and whether a rival took the slot you expected. Then track which sources feed those competitors at domain and page level, sorted by category: competitor-owned pages, editorial reviews, comparison pages, review marketplaces, forums, Reddit, YouTube, and news.

The useful output isn't "Competitor A has more citations." It's:

> Competitor A appears in 8 of 10 Claude runs for the enterprise use-case prompt, mostly through two comparison pages and one review; seoforgpt appears in 2 of 10 and has no cited page for that use case.

Now you know what to build. That gap-to-content loop is the whole reason seoforgpt pairs its tracking with content and outreach agents. Once you can see the brands winning AI discovery and the pages doing it, closing the gap becomes a work order instead of a guess.

Two collection methods exist, and they aren't interchangeable. Active prompt testing sends your own controlled prompts, which gives exact, reproducible, engine-comparable results but may not match what real users ask. User-panel data captures real queries, which feels closer to demand but carries sampling bias and is hard to reproduce. Ask any vendor which they use, or whether it's a hybrid. If they won't say, that's your answer.

One more, since it comes up constantly: the way you upload internal context to a tool matters. seoforgpt's Additional Knowledge workflow accepts DOCX, TXT, Markdown, and pasted text. It does not support PDF, and it does not support PowerPoint files, so convert those before you expect them parsed.

Is a brand citation the same as a good ranking? No, and treating it that way is the most common mistake I see. A citation can support criticism, and a top-ranked mention can lack any source link. Track mention, recommendation, and citation as separate metrics.

How many times should a prompt be run before I trust the result? More than once, always. Given how rarely AI answers repeat, single runs are noise. Report frequency across runs and mark brands as stable, intermittent, or transient.

Can I just check one engine to save time? You can, but you'll get a partial picture. The four engines cite different pages and recommend different brands, so a per-engine report is where the actionable gaps live.

Does blocking or allowing a crawler guarantee citations? No. Allowing `OAI-SearchBot` makes you eligible for ChatGPT Search; `Google-Extended` governs certain Gemini uses. Neither forces an engine to retrieve, cite, or recommend your page.