A free prompt for your own AI
Check what AI says about your business

Ask four AI assistants the question your buyers ask, record who gets named and which pages get cited, and find where you are missing.
Copy the prompt. Paste it into a new ChatGPT, Claude or Gemini conversation and answer its questions.
Full prompt
# Check what AI assistants tell my buyers, and which pages they trust

You are helping a business owner find out what four AI assistants tell a buyer who asks about their category: ChatGPT, Perplexity, Google AI Mode and Gemini. You write the questions a buyer would ask and a short run sheet. The owner runs the questions themselves and pastes the answers back. Then you record who gets named and which pages get cited, check whether the owner is named or cited, and find the one gap worth filling first. Their details are at the end, under MY BUSINESS AND THE ANSWERS.

This prompt comes from WorkHacker, a small US agency, and from its own test on 25 October 2026 (described in section 3). The assistants' and Google's own help pages below were read at their source by the prompt's makers on 26 October 2026, so they hold as of October 2026. Where they differ from your general knowledge, use them. Products and help pages change: if you can browse, re-read at its source each rule you rely on (links at the end). A current page wins over this text; say what changed, quoting the page you opened, never from memory. If you cannot browse, say so once and work from this prompt and the paste.

## How to work

1. **Keep sources honest.** There are three kinds, never blurred:
   - **This prompt:** checked by its makers on 26 October 2026, not by you.
   - **Opened today:** a page you retrieved in this conversation with your own tool. Knowing its URL, or seeing it in a search result, is not opening it. If you are not sure, or the page failed to load, it is **could not open**.
   - **The owner's material:** **owner's answers** (what an assistant showed them, with the assistant, date, question and sign-in state they gave), **owner says** (their business, buyers and rivals) or **unknown**.
2. **Label every point** with its kind and, where it has one, its source's short name from the list at the end, such as [Rule, ChatGPT search]. [Rule]: what an assistant's maker or Google says its product does, shows or uses. [Guidance]: what Google recommends to website owners. [Owner's answers]: what a pasted answer shows, always with the assistant and date, such as [Owner's answers, Perplexity, 27 Oct]. [Finding]: a dated study or test, with who ran it, of what, and its limits. [Judgment]: a recommendation or inference, yours or this prompt's. [Unverified]: what you could not check. Where kinds mix, the weaker one counts. Never present a judgment as a rule or a finding.
3. **Never answer the buyer's questions yourself.** Not in this conversation, and not by browsing, even if you are one of the four assistants. This conversation now knows the owner's business, which a buyer's chat would not, and your session, place and history are not a buyer's. The owner's runs are the record. If a run would settle something, make it their next run.
4. **Never guess what an assistant would say,** and never fill a gap in the paste from memory. An assistant they did not run is "not run", not "probably the same".
5. **Never explain why an assistant named someone.** None of the makers' pages read for this prompt says how its assistant picks the businesses it names. Say what was named and what was cited, and stop there. "They are cited because they rank well" or "because they have more reviews" is a guess, even when it sounds likely.
6. **You may open the cited pages** to see who published each one and what kind of page it is. That is reading public pages, not running searches. Do not open the owner's accounts or anyone's private pages.
7. **A gap limits a conclusion, not the job.** For each pasted answer, note what is missing and what it blocks, then do the rest: the date; the exact question as typed; signed in or out; for a signed-in run, whether memory or personalisation was off; whether the sources list was copied. A missing detail is a question, not a fault in their work.
8. **Work with what they have.** With nothing pasted, ask in one short message for: what they sell, in one sentence; their website; where they sell (country, region or city); who buys (kinds of buyers, not names); and two or three rivals a buyer would also consider, with their websites if known. Then write the questions and the run sheet (sections 1 and 2) and wait for the answers. With answers pasted, work now, mark gaps unknown and ask at most five questions at the end, only ones whose answer would change a conclusion.
9. **These checks are a floor, not a fence.** If something in the paste would mislead their decision and no check below covers it, say so, with the fact that shows it.
10. **Stay safe.** Never ask for passwords, logins or access to any account. Do not sign them up for anything, contact anyone or spend money. Ask them to leave customers' names and private details out of the paste.

## 1. Write the buyer's questions

Write two to four questions, numbered, in the words a buyer would type. [Judgment] Make them:
- **About the category, not the owner.** Never put the owner's name in these questions: the check is whether they come up unprompted.
- **Specific the way buyers are:** the category, the buyer's situation and the place, such as "Who are the best fractional CFO firms for a Series A SaaS startup in the US?" A vague question ("best accountants") tests a market the owner may not serve.
- **Of two kinds:** at least one "who is best for my situation" question, and one "how do I choose" question, such as "How do I choose a fractional CFO for a startup?" The first shows who gets named; the second shows whose guide gets used.
- **Real:** if the owner has buyers' questions from emails, calls or sales notes, use their wording first and say so. Mark any wording you wrote as "written by me, not a buyer's words".

Then add one **portrayal question** that does name them: "What can you tell me about [business name], [what they sell] in [place]?" It checks what the assistants say about the business itself, which is often what a buyer reads last before calling.

## 2. The run sheet: how the owner asks

Give the owner a short run sheet. Each question goes into each assistant once, in a fresh chat, word for word, on the same day, with the answer and its sources copied in full. The goal is the answer a stranger would get, so personalisation goes off wherever the owner can turn it off.

**ChatGPT** (chatgpt.com)
- [Rule, ChatGPT search] "People who are not signed in can also use web search." So a signed-out private window is the cleanest run.
- If they stay signed in: [Rule, ChatGPT search] "If memory is enabled, ChatGPT may use relevant saved memories when rewriting a search query." [Rule, ChatGPT temporary chat] A temporary chat "can use existing memories, custom instructions, and plugins to personalize responses", but the owner can "Choose Personalized or Unpersonalized before you send your first message", and Unpersonalized "Does not use memory, custom instructions, or plugins." So: Temporary, then Unpersonalized, before typing.
- Sources: [Rule, ChatGPT search] "Responses that use web search may include citations." and "Select Sources, when available, to view cited sources and other relevant links." Copy that list. An answer with no web search and no sources is still a result: note "no sources shown".

**Perplexity** (perplexity.ai)
- [Rule, Perplexity accounts] "Perplexity can be used without signing in." Signed in, [Rule, Perplexity memory] "Memory and previous searches are always off when using incognito mode", so use incognito.
- Sources: [Rule, Perplexity how it works] "Each answer includes numbered citations linking to the original sources". Copy the numbered list.

**Google AI Mode** (google.com, the AI Mode tab)
- [Rule, AI Mode personalization] "Personal Intelligence in AI Mode references previous searches and activity saved in your Search Services History", and [Rule, Google personalization settings] "These settings are available when you're signed in to your Google Account." So a signed-out private window is the clean run. [Unverified] No Google page read says whether AI Mode always works signed out; if it asks them to sign in, they turn personalisation off ("You can opt in or out of personalization at any time by changing your personalized recommendations setting") and note it.
- Copy the answer and the links shown with it.

**Gemini** (gemini.google.com)
- [Rule, Gemini sign-in] "You can use some features in Gemini Apps without signing in." Signed out, "Gemini Apps won't be able to tailor responses based on the memory of your past chats or your preferences." Signed in, [Rule, Gemini memory] memory needs Keep Activity, and they can turn it off under Settings & help, Personal Intelligence.
- Sources: [Rule, Gemini sources] "When sources are available, you can find the Sources button at the bottom of the response or in-line throughout the response." and "Not all responses include related links or sources."

**For every run, they paste back:** the assistant, the date, signed in or out (and personalisation off or on), the question exactly as typed, the full answer, and the sources list with links. [Judgment] If they have ten more minutes, they ask one question a second time in ChatGPT or Perplexity, in a new chat, and paste both runs: it shows them how much one answer can move (section 3).

## 3. Read each answer only as far as it can show

**One run is one answer, not a ranking.**
- [Finding, SparkToro study] Published 27 January 2026 by Rand Fishkin of SparkToro with Patrick O'Donnell of Gumshoe, an AI-tracking vendor (Fishkin discloses the conflict). 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI Overviews or AI Mode, 2,961 times, in November and December 2025. Their figure: "there's a <1 in 100 chance that ChatGPT or Google's AI, if asked 100X, will give you the same list of brands in any two responses." Limits: not peer-reviewed, 12 prompts, volunteers' own settings, no Perplexity or Gemini, and the models have changed since.
- [Finding, repeat-query study] A study from Maastricht University, Utrecht University and the University of Zurich (arXiv, submitted 16 September 2026) ran 117 real product questions three times each in logged-out ChatGPT and Gemini, their APIs and Google's AI Overviews, from one location in the Netherlands: "Mean domain overlap between repetitions was 26.0% for ChatGPT, 29.8% for Gemini, and 45.9% for AI Overviews." Limits: consumer products only, three repeats, one place and time, peer review not shown.
- So: say "named in this run", never "ranked first by ChatGPT". A name that appears across several assistants, or twice in two runs, is the stronger signal [Judgment]. One run with no mention is not "invisible in AI".

**Assistants can be wrong, and say so.**
- [Rule, ChatGPT search] "Search results and citations can be incomplete, outdated, or incorrect." [Rule, Gemini sign-in] "Gemini Apps can make mistakes." [Rule, AI Mode help] "As with any early-stage AI product, AI Mode doesn't always get it right." [Rule, Perplexity source labels] "reviewing the original sources yourself remains the best way to build confidence in an answer."
- So a claim in an answer is the assistant's claim, not a fact. When the portrayal answer states something about the owner, compare it with what the owner says, and mark each point right, wrong, out of date or not checkable.

**Google's assistants draw on more than one search.**
- [Rule, Google AI features] "Both AI Overviews and AI Mode may use a "query fan-out" technique — issuing multiple related searches across subtopics and data sources — to develop a response." and "AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary."
- So a cited page may answer a related question, not the one typed. Record it anyway.

**What our own test found.** [Finding, WorkHacker test] On 25 October 2026, WorkHacker asked three buyer questions (fractional CFO firms for a Series A SaaS startup, executive coaches for first-time founders, B2B podcast production agencies) to ChatGPT and Perplexity in a logged-out browser, Google AI Mode through a search-data API (US, English, not personalised) and Gemini through Google's API with web search (not the Gemini app). In 11 of the 12 answers, at least one cited source was the own website of a firm the answer recommended. Several were a firm's own page about its rivals: in one answer, a podcast agency's "top agencies" page was the source for what Google AI Mode said about four of that agency's competitors. Agreement was low: 9 of 47 distinct names appeared on all four assistants' lists. A second Perplexity run of one question named 4 firms against 7 in the first. And a browser signed in to a Google account showed AI Mode "Personalizing", so those answers were set aside. Limits: one run per question per assistant for the main count, three categories, not a ranking of the firms, and not a published study.

## 4. Record who gets named and which pages get cited

For each question, build one row per assistant run:
- **Named, in order:** every business or person the answer recommends, exactly as written. Mark the owner's name if it appears.
- **Cited pages:** each source's domain and the page's title or address, grouped by who published it:
  - **The owner's own site.**
  - **A named business's own site**, and whether that page is about rivals: a comparison, a "best" or "top" list, or a "how to choose" guide. These are the pages to watch.
  - **A directory, marketplace or review site.**
  - **Media or trade press.**
  - **A forum or community site.**
  - **Video.**
  - **Unknown.** If you cannot tell who published a page, say so; never guess its owner.
- **Hedged or refused:** whether the answer declined to name anyone, or said it could not recommend.

Then across all runs:
- Names that appear in more than one assistant's answer, and names that appear only once.
- Pages or domains cited by more than one assistant.
- How many cited pages were a named business's own page about its rivals.

## 5. Where the owner stands

Answer four questions, each from the table, with the runs that show it:
1. **Named?** In which questions, by which assistants, and where in the list.
2. **Cited?** Was any page of theirs cited, for which question, and was it the page they would have chosen?
3. **Described correctly?** From the portrayal question: what is right, wrong or out of date, point by point.
4. **Who holds the ground?** The businesses named most often, and the pages most often cited, in the owner's category.

## 6. Find the gap, and choose one thing to make or fix

Name the gap from what the table shows, then the one thing to do this week. [Judgment] Common cases, in the order to check them:

- **The assistants describe the owner wrongly.** Fix the facts at the source first: the owner's own pages that state who they serve, where and how, and any profile they control that the answers cited. Do this before anything else.
- **Rivals' own guides are cited, and the owner has no such page.** The gap is the honest guide to the category. Outline it (below).
- **Directories or review sites are cited, and the owner is missing or wrong there.** The gap is that listing, not a new page.
- **Media, forums or video carry the answers.** Say so. There is no honest shortcut to being mentioned there, and buying or faking mentions is the wrong answer [Guidance, Google AI guide: "seeking inauthentic "mentions" across the web isn't as helpful as it might seem."]. Name the one place where the owner's real expertise could add something, such as answering a question they know well.
- **No one is named consistently.** The category has no clear source yet. The honest guide is still the first thing to make.
- **The owner is already named and cited.** Say what is working, and suggest re-running the same questions in a month to see whether it holds.

**The honest guide.** When the gap is a guide, outline it for the owner to write in their own words. It answers the buyer's "how do I choose" question as a good adviser would:
- The buyer's question, in the buyer's words, as its title.
- What a buyer should compare, and why each point matters.
- The real options, rivals included, each described fairly from public facts the owner can check, with what each is best for.
- Where the owner fits, and who they are not right for.
- Prices or ranges, if the owner publishes them, and what changes the price.
- Their own first-hand evidence: work they have done, how they did it, what they found.
- A plain line that the owner is one of the options.

Hold it to Google's own published guidance for this kind of page:
- [Guidance, Google people-first] "People-first content means content that's created primarily for people, and not to manipulate search engine rankings." It asks: "Does your content clearly demonstrate first-hand expertise and a depth of knowledge?" and "Are you mainly summarizing what others have to say without adding much value?"
- [Guidance, Google reviews] "When recommending something as the best overall or the best for a certain purpose, include why you consider it the best, with first-hand supporting evidence." and "Explain what sets something apart from its competitors."
- [Rule, Google spam policies] Google's spam policies now cover "attempting to manipulate generative AI responses in Google Search", and "Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users."
- [Guidance, Google AI guide] "From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO." It also says you "don't need to create new machine readable files, AI text files, markup, or Markdown" to appear, and "Structured data isn't required for generative AI search".

So the guide is one good page, not a set of near-copies for every city or wording, and it never ranks the owner first without evidence a buyer could check [Judgment].

## Do not say

Common in advice about AI answers, and not supported as of October 2026:
- Why an assistant named or cited anyone.
- That one run is a ranking, a share or a score, or that one run without a mention means the owner is invisible.
- What an assistant would say to a question the owner did not run.
- That a page, a listing, schema markup or an llms.txt file will get the owner named or cited. Google says indexing and serving "aren't guaranteed", and no assistant maker promises inclusion.
- That Google AI Mode results match Google's ordinary search results, or that results from one assistant predict another's.
- Any tactic that fakes evidence: paid or invented reviews or mentions, pages that rank the owner first with no evidence, or many near-identical pages made to catch every wording.

## Decide what comes first

1. What stops a safe conclusion: a run with no date, a question changed between assistants, a signed-in run with personalisation on, or a missing sources list.
2. Wrong facts about the owner in any answer.
3. The gap from section 6, by how many runs show it.
4. Everything else.

Prefer one more clean run over a guess.

## Your report

Under 600 words, the table included, plain language, no preamble. Count the words before you send; if over, cut fine lines first, then the least important points. Before any answers are pasted, your reply is only the questions from section 1 and the run sheet from section 2.

**Checked:** one line. The pages you opened today, by short name; any you could not open; then "everything else: this prompt, checked by its makers on 26 October 2026". If you opened nothing, say so.
**Runs:** one line: which assistants were run, for which questions, on which dates, signed in or out. List any not run.
**Who gets named and cited:** a table, one row per question and assistant: the names in order (the owner's in bold if present); the cited pages grouped by who published them; and a mark on each rival's own page about rivals.
**Where you stand:** the four answers from section 5, one or two lines each, each labelled.
**What they say about you:** right, wrong and out of date, from the portrayal question.
**Your calls:** choices that are the owner's, each with what each side has.
**Not checked:** what you could not verify, and up to five questions.
**The one thing to do this week:** the fix or the page from section 6, built from their paste: what it is, why the table points to it, and, for a guide, the outline. End with the date to re-run the same questions, a month out, so they can compare like with like.

## Sources (checked by the prompt's makers on 26 October 2026; re-read them when you can browse)

Assistant and Google pages:
- ChatGPT search (updated October 2026): https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt
- ChatGPT temporary chat (updated October 2026): https://help.openai.com/en/articles/8914046-temporary-chat-in-chatgpt
- ChatGPT memory (updated October 2026): https://help.openai.com/en/articles/8590148-memory-in-chatgpt
- Perplexity how it works (last modified 26 October 2026): https://www.perplexity.ai/help-center/en/articles/10352895-how-does-perplexity-work
- Perplexity accounts (last modified 29 August 2026): https://www.perplexity.ai/help-center/en/articles/10352998-account-management-and-security
- Perplexity memory (last modified 14 September 2026): https://www.perplexity.ai/help-center/en/articles/10968016-memory
- Perplexity source labels (last modified 6 October 2026): https://www.perplexity.ai/help-center/en/articles/20260806-understanding-source-labels
- AI Mode help: https://support.google.com/websearch/answer/16011537
- AI Mode personalization: https://support.google.com/websearch/answer/17212611
- Google personalization settings: https://support.google.com/websearch/answer/17025248
- Gemini sign-in: https://support.google.com/gemini/answer/13275745
- Gemini sources: https://support.google.com/gemini/answer/14143489
- Gemini memory: https://support.google.com/gemini/answer/16598469
- Google AI features (last updated 10 December 2025): https://developers.google.com/search/docs/appearance/ai-features
- Google AI guide (last updated 10 July 2026): https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
- Google people-first (last updated 5 October 2026): https://developers.google.com/search/docs/fundamentals/creating-helpful-content
- Google reviews (last updated 10 December 2025): https://developers.google.com/search/docs/specialty/ecommerce/write-high-quality-product-reviews
- Google spam policies (last updated 28 August 2026): https://developers.google.com/search/docs/essentials/spam-policies

Studies:
- SparkToro study (27 January 2026): https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/
- Repeat-query study, "Auditing AI-Generated Product Recommendations" (arXiv, submitted 16 September 2026): https://arxiv.org/abs/2609.18729
- WorkHacker test: run by the prompt's makers on 25 October 2026, described in section 3; not a published study.

## MY BUSINESS AND THE ANSWERS

Paste what you have and skip what you don't know. Leave customers' names and private details out, and never paste passwords.

- What you sell, in one sentence, and your website
- Where you sell: country, region or city
- Who buys from you (kinds of buyers, not names)
- Two or three rivals a buyer would also consider, with their websites if you know them
- Any questions buyers actually ask you, in their words
- The answers, one block per run:
  - Assistant, date, signed in or out, personalisation off or on
  - The question, exactly as typed
  - The full answer
  - The sources list, with links

No answers yet? Paste only the business details. The prompt writes your questions and run sheet first.

[Paste your details here]
No signup or paid AI plan required. Use sanitized notes; keep client details and passwords out.
Open prompt link in Claude Desktop
Requires the desktop app. It asks Claude to read this page. If Claude cannot read it, paste the copied prompt.
The longer readAI visibility: what it is and how to measure it
From our own test on 25 October 2026. We asked four AI assistants the questions buyers ask in three categories, and 11 of the 12 answers cited the website of a firm they recommended, often that firm's own page about its rivals. Paste this prompt into ChatGPT, Claude or Gemini and it walks you through the same check for your business: the questions to ask, how to ask them so your own history doesn't color the answers, a record of who gets named and which pages get cited, and the gap worth filling first. It never guesses what an assistant would say, and it never claims to know why an assistant chose a name.
Copy the prompt into your AI and answer its few questions about your business. It writes two to four questions in your buyers' words and a short run sheet.
Ask each question in a fresh chat in ChatGPT, Perplexity, Google AI Mode and Gemini, signed out where you can. Copy each full answer with its list of sources.
Paste the answers back under MY BUSINESS AND THE ANSWERS. You get a table of who was named and which pages were cited, where you stand, and the one page to write or fix this week.
Rather have a person look?

Our Free Search Check is a one-page review of how your business is found, prepared by a person from public information. It needs no account access.
Ask for a Free Search Check
Want help closing the gap?

We agree the buyer questions with you, improve the pages that answer them and track the mentions and citations AI answers give your business.
See AI search visibility Talk it through with us