How we observe AI answers.
We ask the OpenAI, Google Gemini, Anthropic Claude and Perplexity search APIs set questions. Here is what we record. Each protocol keeps its own date; results are never mixed.
What a result shows · Report preview · Done-for-you delivery details
AI visibility check (from October 20, 2026).
- Questions. Up to 10 in a first check: without your brand name (a buyer is choosing) or with it. Each notes its source: you, or a template or model draft. Drafts stay marked until you confirm them and never count as search demand. An edited question is a new version, never compared with the old one.
- Collection. Since October 28, 2026: each question twice on each of four AIs, each with its own search: the OpenAI Responses API with web search, the Gemini API with Grounding with Google Search, the Claude API with web search, and the Perplexity Sonar API. Earlier checks asked three times on OpenAI and Gemini only, and are not compared directly. Each answer saves its date, AI, requested and reported model, market, language, search settings, whether a search ran, cited URLs, full text and status.
- Who is asking. Each AI gets the same note: who is asking (your buyer, or a default for your business type), market and language. OpenAI, Claude and Perplexity also get a rough location; Gemini’s search tool can’t. A market or language we can’t measure is marked unsupported, never swapped.
- Status. Waiting, succeeded, failed or not collected. Retries are limited and add no answers. Failed answers are left out of rates, shown in coverage and never counted as “not mentioned”.
- Signals. Mentions, links to your domain, rivals and cited sources are counted apart. “We found nothing about you” is not a mention. A person judges recommendations from the exact sentence. Answers that differ from your confirmed facts show side by side as possible errors.
- Comparisons. Only when questions, AIs, requested models, repeats, market, language and search settings all match. Otherwise: “not comparable”, with what differs. A new model build is flagged.
- Usage. Each check saves planned answers, calls, failed attempts and estimated API cost.
Rules 3 and 5 below apply to these checks too.
Done-for-you pilot (October 19, 2026).
The rules below were written for the human-run pilot and still apply to it.
1. Pick the questions first.
Up to 20 questions, one product, one language, first from real sales objections, enquiries and FAQ. Each saves its source type, a reference, the date checked and whether your team confirmed it. Model or template drafts are marked a guess, not proof of demand. No private material or personal data goes to an AI.
Brand-name questions stay apart from choosing questions. We never swap questions for nicer answers. If questions or settings change, before and after are not compared.
2. Ask more than once, and keep the answers.
Each question twice on each of the four search APIs, on three dates: two separate “before” dates and one agreed date after release. That is at most 20 × 2 × 4 × 3 = 480 planned answers. Search is on, but not every answer searches or cites; Gemini decides when to search.
These are API answers: not the ChatGPT, Gemini, Claude or Perplexity apps, nor Google AI Overviews or AI Mode. Our scanner supports a few buyer countries and one buyer role; otherwise we agree a manual method in writing and report it apart.
- Each run saves date, buyer region, questions, settings, reported model names, answers and citations.
- Reports show successful against planned answers. Failed or empty answers stay visible but out of rates.
- A cached answer is not a new observation.
Two repeats show variation, not statistical significance. Models and search can change under the same name.
3. Count each signal on its own.
- Brand mention rate: successful discovery/alternatives answers naming your brand or an agreed alias, over all successful answers in that group, per AI. A mention can be neutral or negative.
- Positive recommendations: answers a person found clearly recommend you, with the reviewer, the exact quote and how many were reviewed.
- Own-domain citations: answers citing your agreed domain or subdomains, same denominator. Not a visit.
- Visits and pipeline: optional GA4/CRM totals you share, with source and period. Missing means “not supplied”, not zero.
- Enquiries: ones your system really received that meet your approved rules, minus spam, duplicates and tests. An email-link click is not one. Events, ad numbers and enquiries are never summed.
None of this is market share, search volume or a count of buyers.
4. Compare only matching runs.
Same questions, buyer region, settings and reported models, and at least 95% of planned discovery/alternatives answers succeeding in both runs. Otherwise the report shows no change and says why. Changes are in percentage points from matched answers. The 95% rule is a coverage check, not a confidence level or significance test. A matching change still does not prove a page edit caused it.
5. Mark real ChatGPT and Gemini app checks apart.
If the scope adds an app check, we save the first answer to a set question in a new chat, with date, exact prompt, shown model or mode, search setting and citations (or “not shown”). App answers vary with history, settings, account, location and search. A US question typed from Korea is not a US session, and one missing mention doesn’t mean buyers can’t see you.
We keep full answers and screenshots, record hits and misses, never re-ask for a better sales line, and keep app checks out of API counts.
6. Use the proof to pick the work.
A missing mention alone doesn’t prove your site is broken. We read the page, then set side by side the buyer’s question, your confirmed facts and what the page misses or gets wrong. A page is picked when answers point to it and it lacks a fact buyers asked for. Changes add only confirmed facts; missing facts are asked for.
Each job saves URL, proof, change, approver and exact approved version. Done means the live version matches and agreed tests pass. We ask once more on the agreed date. A difference doesn’t prove the change caused it; business effects stay “not assessed” without enough data. Ranking, traffic, revenue and AI recommendation increases are not guaranteed.
About the September 2026 research.
That exploratory sample used a different protocol. Its answers don’t keep every model and request setting, and questions repeated across companies and AIs. Its rates use only choosing questions (discovery and alternatives), not company-fact or head-to-head ones. See the done-for-you delivery details for what each job includes.
Legacy weekly protocol.
Existing analysis orders were written against a weekly protocol of 50 questions on the OpenAI, Google Gemini and Perplexity APIs, three repeats each (450 planned answers). That protocol applies only to those written orders. Perplexity joined the AI visibility check protocol on October 28, 2026, and the improvement pilot uses it too.