How do I check whether ChatGPT, Claude and Perplexity recommend my site?
Ask the same 14 questions in each assistant, logged out and in a fresh chat: 12 that your customers would ask without your brand name, plus “What is [brand]?” and “What is [brand] ([domain])?”. For each answer, write down whether you were named, recommended first and cited, then ask the same questions again 14 days later. It costs nothing but your time, and it tells you more than any single answer, because answers change from run to run and API answers differ from what the apps show.
The short answer
Ask the same 14 questions in each assistant, logged out and in a fresh chat, and log three things for every answer: were you named, were you recommended first, and was one of your pages cited. Ask again 14 days later. Count how often you’re named across the whole set, not your position in one answer: one answer is a sample, not a rank.
Step 1: write 12 + 2 questions
Twelve questions your customers would ask, without your brand name, and two diagnostics that name you. This is the set from our book, also used by the Katman method:
| # | Type | Template | What it tells you |
|---|---|---|---|
| 1–3 | Situation | “[Situation] but [obstacle]. Is there a site that [result]?” | The type small sites win |
| 4–5 | Tool | “Which tool does [job]?” | Short and generic |
| 6–7 | “Best X” | “Best [category] 2026” | Big brands dominate |
| 8–9 | Alternative | “[Big competitor] alternatives with [feature]” | List articles |
| 10–11 | Language or market | “Best site for [job] in [language]” | Local competitors |
| 12 | Problem | “How do I fix [problem]?” | Usually names no tools |
| T1–T2 | Diagnostic | “What is [brand]?” and “What is [brand] ([domain])?” | Confusion; is your domain known? |
Write the twelve the way customers talk: take them from sales calls, support emails and reviews. Situation questions matter most for a small site. On our sample site, ChatGPT named the site first for a question describing the user’s situation, and didn’t name it for “best [category] 2026”, where it listed four big brands (own measurement, 26–28 September 2026).
Step 2: ask in a clean session
- Open a private window and stay logged out. If an assistant won’t answer without an account, start a new chat that doesn’t draw on your history.
- Turn web search on.
- Ask each question in a new chat, word for word, in ChatGPT, Claude and Perplexity; add Gemini and Copilot if your customers use them.
- If a logged-out session stops you after a few questions, open a new one.
- Save every answer with its sources and the date, as a screenshot or pasted text.
Why so strict: on 4 September 2026, a signed-in ChatGPT account placed one of our own sites second for a buyer question, while a temporary chat, logged out, didn’t name it at all (own measurement). Your account knows your history; your customers’ accounts don’t. Fourteen questions in three assistants make 42 answers per round.
Step 3: record named, recommended and cited
Write one row per question and assistant, with the date, the assistant and the question number, and these four columns:
| Column | What to write |
|---|---|
| Named | Yes, if your brand appears anywhere in the answer |
| Recommended | Yes, if the answer suggests you as an option; note your position and whether you came first |
| Cited | The URL of your page if it’s among the sources, otherwise “none” |
| Also named | The other brands, and the domains the answer cites |
Keep named, recommended and cited separate, and don’t merge them into one score. For T1 and T2, also check that the description is right: is it your product, or something with a similar name?
The “also named” column becomes your competitor list, and the cited domains your source list: the pages assistants read when they answer your customers.
Step 4: ask again after 14 days
Same questions, same assistants, same conditions. Compare how many of the 14 questions named you, recommended you and cited you, and note what you changed in between. Checking more often mostly measures noise: answers vary between runs, and a change on your site takes time to be crawled.
Why one answer misleads
- Answers are unstable. Ask ChatGPT or Google’s AI the same question 100 times and there’s less than a 1 in 100 chance of getting the same brand list twice (SparkToro; third-party study). Frequency across many runs means more than any single rank.
- API answers differ from app answers. In a study of 1,000 prompts, brands named through the APIs overlapped with the apps’ answers by only about 15–32% (Surfer, updated 25 September 2026; third-party study, from a vendor that sells app-based tracking). In our own small comparison in September 2026, the API and logged-out ChatGPT agreed on whether the brand was present in 3 of 3 cases, but named the same brand first in only 1 of 3 (own measurement). Read presence as fairly reliable and rank as unreliable.
- Branded questions flatter you. On 3 September 2026, on another site we run, the assistants recognised the brand in 19 of 22 branded questions and named it in none of about 30 unbranded buyer questions (own measurement).
- Each assistant is different. Our sample site was named in 0 of 12 buyer questions in Perplexity in its first week, while ChatGPT sent it 333 visitors (own measurement, 21–27 September 2026). Check each assistant separately.
Two official reports count real citations and clicks instead of a sample: Search Console’s Generative AI performance report for Google’s AI features (Google AI optimization guide) and the AI Performance report in Bing Webmaster Tools for Copilot (Bing, 10 February 2026). Count ChatGPT’s visitors in your own analytics, by a chatgpt.com referrer or utm_source=chatgpt.com in the link: 79% of our sample site’s ChatGPT visitors sent no referrer (own measurement).
The API option: a Katman visibility run
If asking 42 or more questions by hand every two weeks takes too long, Katman can ask them for you. katman_visibility_run sends your 12 + 2 questions to the models behind ChatGPT, Claude, Gemini, Perplexity and Grok through OpenRouter, with live web search, on your own OpenRouter key. It records what you would: named, recommended first and cited, plus the other brands and the domains cited.
- Cost: about $2.4 per run on the default panel, with a range of $1.2–3.7, paid by you to OpenRouter. What drives the cost is the search results the models read, not the length of your questions.
- No surprises: by default a run is a dry run that lists every call, the estimate from OpenRouter’s live prices and your key’s remaining credit. Nothing is spent until you agree.
- Capped at $3 by default: a run won’t start if the estimate is above your cap, and it stops starting new calls when it reaches it. Raise
max_cost_usd, use theprobepanel (ChatGPT and Perplexity only, about $1.8), or ask fewer questions. - Plan: the tool comes with Audit Week ($3 once, for 7 days) or Katman Pro ($29 a month at the founding price for the first 100 customers); the OpenRouter cost is separate. Audit Week ends before day 14, so the second run needs Pro, or your manual table. With Pro,
katman_setup_monitoringcan schedule the runs in GitHub Actions. - Results stay in your project: a table in
.katman/measurements.md, with the changes since the previous run and the date of the next one.
API answers are a proxy for the apps, so confirm your most important questions by hand. katman_manual_check gives you the exact questions and records the answers you paste in, with no API cost; Katman never drives the ChatGPT, Claude or Perplexity apps itself. How runs work is on visibility runs. If you need a hosted dashboard for a team instead, see Which AI visibility tool fits your situation?.
What to do with the results
| What you see | Likely cause | Where to go |
|---|---|---|
| Not named anywhere, not even for “What is [brand]?” | AI bots are blocked, or your text isn’t in the HTML | My vibe-coded site is invisible to ChatGPT, Cloudflare is blocking AI crawlers, the AI crawler checker |
| Named, but described as something else | The assistant confuses you with another product | Cause 4 on Why doesn’t ChatGPT recommend my brand? |
| Known by name, never recommended for buyer questions | No quotable first sentence, no page for the customer’s situation | Situation pages in the Katman method |
| Recommended for situations, not for “best X” | Normal for small sites: list articles and big brands win those | An honest comparison guide and mentions elsewhere (causes 6 and 7 on the same page) |
| Competitors named, their pages cited | These are the pages assistants read for your customers’ questions | Ask the assistant which articles it used, and send each author a short, factual note |
When you fix something with your AI coding tool, give it measurable jobs rather than “optimize SEO”: see what to ask Cursor instead. Note every change with its date, so the next round shows what it did. No one, including us, can guarantee a place in an AI answer, and Google says no third-party tool has access to its internal ranking or AI systems (Google AI optimization guide).
Sources: the SparkToro and Surfer studies linked inline (third-party); Google AI optimization guide; Bing Webmaster blog; all checked 30 September 2026; our own measurements as dated. Written by the Katman team, Rnify Limited.
Who this is not for
- Teams that need a hosted dashboard with seats, client reports and daily tracking: compare the tools built for that.
- Anyone looking for one AI visibility score or a single rank: answers vary too much for one number to mean much.
- Sites whose pages AI bots can’t read yet: fix that first; the questions can wait.
Frequently asked questions
Can I check this without paying for a tool?
Yes. Ask the 14 questions by hand in each assistant, logged out, record whether you were named, recommended and cited, and repeat after 14 days. It costs only your time.
ChatGPT recommends my site when I ask. Why don’t my customers see it?
Because your signed-in account remembers you. On 4 September 2026, a signed-in ChatGPT account placed one of our sites second, while a logged-out temporary chat didn’t name it at all (own measurement).
Is asking “What is [brand]?” a good test?
No. It shows whether an assistant knows you, not whether it recommends you. On 3 September 2026, assistants recognised one of our brands in 19 of 22 branded questions and named it in none of about 30 unbranded buyer questions (own measurement).
How often should I check?
Every 14 days, with the same questions and the same assistants, is enough for most small sites. Checking more often mostly measures noise.
Do API-based tools show what ChatGPT users see?
Not exactly. In a Surfer study updated on 25 September 2026, brands named through the APIs overlapped with the apps’ answers by only about 15–32%; treat presence as fairly reliable and rank as unreliable.
What does a Katman visibility run cost?
About $2.4 per run on the default panel (range $1.2–3.7), paid by you to OpenRouter with your own key, shown before anything runs and capped at $3 unless you raise the cap. The tool itself comes with Audit Week ($3) or Katman Pro.
Can any tool guarantee that I’ll be recommended?
No. Answers change from run to run, and Google says no third-party tool has access to its internal ranking or AI systems.