You give us a domain and a category
In your buyer's words, not your company's. Something like "ai visibility tracking for agencies," not "unbound geo." We derive suggestions from your domain, but you write the category.
We turn that into real questions and ask live engines
Shortlist, comparison, problem and local questions, run against ChatGPT, Claude, Gemini and Perplexity with web search turned on. One call at a time, never in parallel, never simulated.
We store every answer
The full raw response and the final message text, saved with the model, the cost and the timestamp. Nothing is recomputed from memory later; it is read back from what actually happened.
We read your brand out of the stored text
Detection runs on saved text, never live. It scores how confident the match is, how prominent the mention is, and whether it is a recommendation or a dismissal, then reports the verdict, the findings and the diagnosis.
A run across 15 prompts and 4 engines plans 55 calls, costs about $4.01, and stores 51 answers with 460 cited URLs across 190 distinct domains. Measured on the first full run we reconciled against provider billing.
Every prompt is written in the buyer's language, never the company's. A buyer does not type your brand name into ChatGPT and ask it to grade you. They type what they actually want: "best growth marketing agencies for AI startups," "odontólogo en Chapinero," "how do I fix low AI visibility." Your name is not in the question, and it should not be, because the point is to see whether the engine surfaces you on its own.
There is exactly one exception: the self_description prompt, where we ask the engine directly what it knows about your company. That prompt exists to catch the engine getting your own story wrong, not to measure share of voice, and it is excluded from that number entirely.
A report is a saved, dated object with its own URL, not a live dashboard that changes underneath you. It carries five things.
the verdict
One sentence. Whether you showed up at all, and where. Not a chart, a sentence you can quote.
the receipt
The actual answer text, the exact prompt, the engine, the model and the date. This is the part no competitor puts on the page: you can read the paragraph that produced the number instead of trusting it.
who got named, in what order
Every brand the engine mentioned, in the order it mentioned them, including competitors nobody told us to track. In one validation run, a brand with a distinctive name appeared in 3 of 4 engines with a weighted share of 12%, while the category leader held 24%, and the scan surfaced a competitor, ClickUp, that nobody had listed.
the cited sources
Every concrete URL the engine read to build its answer, ranked by frequency. This is the press and content target list: the pages that already influence what the engine says, whether you are on them or not.
named findings
Not a score. Specific, evidenced things that happened, covered next.
A percentage is not a finding. "Your share of voice is 12%" does not hurt anyone and does not tell anyone what to do. There are exactly three types of finding, and every one of them carries a proper name and evidence you can open.
A named competitor appears in the answer and you do not. Evidence: the answer text, the competitor's rank or context, and the prompt that produced it.
The engine cited a concrete URL to build its answer, and that page names competitors but not you. Evidence: the cited URL, its domain, and the answer it fed.
On the one prompt that is allowed to name your company, the engine describes you with something wrong, outdated or incomplete. Evidence: the exact sentence, quoted.
We do not average across engines and call it a day. Each one has a different cost, a different citation behavior, and a different failure mode, and the product treats them differently on purpose.
| engine | class | avg cost / call | range |
|---|---|---|---|
| Perplexity | Search-native | $0.0256 | $0.0116 to $0.0376 |
| Claude | Frontier | $0.0631 | $0.0555 to $0.0819 |
| Gemini | Fast | $0.0702 | $0.0104 to $0.1123 |
| ChatGPT | Frontier | $0.0964 | $0.0345 to $0.1189 |
Perplexity is the workhorse
Cheapest of the four by a comfortable margin, and it returns real web sources, typically 11 to 17 of them per answer. Best cost per cited source of the mix, which is why it is the default engine for the free scan and the low tiers.
ChatGPT is the demo engine
The most expensive of the four, and part of what you pay for is a reasoning trace we exclude by design: reasoning models return an internal draft alongside the final message, and that draft can name brands the final answer drops. Only the final message is ever analyzed, so its real cost per usable answer is worse than the sticker price.
Claude is the middle option
Not the cheapest, not the most expensive, and its cost per call does not swing as wildly as Gemini's or ChatGPT's.
Gemini is the weakest fit
About 2.7 times the cost of Perplexity per call. It rejects the country code parameter, so its runs carry no local segmentation, which matters a lot for markets outside English. Its citations arrive as opaque grounding redirects rather than direct URLs, so its source data cannot be resolved into a press target list without extra work. It sits outside the default engine mix and is available as an explicit option.
Measuring is the door. Fixing is the business. Every finding maps to a probable cause and a discipline that addresses it, and when a finding does not trace back to something we can fix, the diagnosis says so instead of manufacturing a reason to sell you something.
| cause | discipline |
|---|---|
| You do not appear in any media the engine reads | PR and earned media |
| Your site does not explain what you do in extractable terms | Content and messaging |
| No external validation: reviews, directories, threads | Partnerships and community |
| A competitor owns the category narrative | Positioning |
| No citable proof points | Case studies |
| Your founders are not visible as authorities | Founder content and KOLs |
| You are absent where buying conversations happen | Business development and events |
Five real questions against a live engine with web search on. No account, about a minute, and you get the answer text, not a score.
Written 30 July 2026.