You asked ChatGPT which businesses could help with a customer’s problem. Maybe yours appeared. Maybe it didn’t. Now you want to know whether Claude, Gemini, and Perplexity give a different answer—and what that difference means.
A ChatGPT check is a useful starting point, but it cannot stand in for every assistant a customer might use. To check whether AI recommends your business across multiple assistants, reuse the same customer question and preserve enough evidence to compare the answers fairly.
This guide shows you how. It builds on our ChatGPT recommendation-check guide, rather than repeating its full walkthrough.
Keep one limit in mind from the start: answers vary with wording, location, time, provider, and settings. One answer is an observation, not a permanent rank.
Why one engine isn’t enough
Different assistants can retrieve different sources and turn similar information into different answers. One might offer a shortlist of businesses. Another might explain how to choose a provider without naming anyone. A third might link your website without suggesting customers hire you.
Those differences matter, but no single engine is “the ranking that matters.” The aim is to spot patterns across a small set of questions your customers would actually ask.
If the term is new to you, AI Visibility for Small Businesses explains the broader idea: whether assistants can find, describe, cite, or recommend your business. Here, we’re focusing on comparing what they actually return.
What you are—and aren’t—measuring
“Tell me about [your business]” tests what an assistant can say about a name you supplied. An unbranded customer question tests whether it brings up your business without that help. Keep those tests separate.
Use these labels when reading each answer:
- Recommended: Your business is presented as a choice for the customer’s need. A shortlist can count; an explicit “we recommend” is not required.
- Mentioned: Your business is named, but not offered as a suitable choice. Read the surrounding wording.
- Cited only: Your website or a page about your business is visibly cited, but the business is not named as a choice or discussed in the answer text.
- Not present: Neither your business nor an identifiable source about it appears in the answer or visible sources.
- Incorrect: The answer gives a wrong business fact, such as the city, services, hours, or operating status.
“Incorrect” can overlap with the other labels. Record “recommended; incorrect service area,” for example, rather than hiding the error behind a positive result. Likewise, record citations separately when a business is both recommended and cited.
The labels above are for your manual comparison. In Honest Visibility’s complete report, the automatic customer-search labels are “Appeared,” “Not named,” or “Unavailable,” with sources and business-accuracy results shown separately. Read the recorded answer to see whether an appearance was a recommendation or just a mention.
This check does not prove what every customer will see, reveal an assistant’s internal selection process, or guarantee future placement.
Prepare one fair customer question
Use an unbranded prompt
Start with this pattern:
Which [type of business] can help with [specific need] in [location or service area]?
Leave out your business name, domain, and owner name. Don’t add a distinctive slogan to steer the answer toward you.
For a current-business check, you can include “Use current web sources where available.” If you use that instruction, include it in the exact prompt you paste into all four assistants.
Prefer a specific need over vague “best”
“Who is the best contractor?” leaves too much unspecified. A clearer question would be:
Which contractors can remodel a small kitchen in Atascadero while the homeowners continue living there?
Other possible questions include exterior cleaning for a stucco house in Santa Maria or bookkeeping for a small construction company in SLO County. These are examples, not measured keywords. Use your real service, customer need, and geography, whether you work on the Central Coast or elsewhere.
Name the location instead of relying on “near me.” For a remote service, describe the actual customer market or service area.
Decide your small prompt set before you start
Choose two or three questions at most: your core service and location, a more specific need, and perhaps a nearby city you genuinely serve. Change one meaningful part at a time.
Save those questions before opening the assistants. Run each unchanged across all four. Don’t keep rewriting or rerunning until you appear; that gives you a favorable screenshot, not a fair comparison.
Run the same check in all four assistants
Use a fresh conversation for each question where possible, without first discussing your business. A fresh thread reduces the influence of that thread’s earlier messages; don’t assume it removes account-level personalization.
If the test needs current local businesses, enable web search or browsing where offered. Check the product’s current settings instead of following an old screenshot. Record the mode or model shown, whether search was enabled, and whether the response visibly used it. Enabled and used are not necessarily the same thing.
Try to run the comparison in one sitting. If one assistant cannot search or fails to answer, record the limitation—not a business omission. Label unknown settings “unknown.”
ChatGPT
Paste the saved question into a fresh conversation. Note whether search was used and any known location or personalization context. Save the full response and visible sources.
For the deeper setup and source-inspection steps, use the existing ChatGPT guide. The extra task here is keeping its question consistent with the other three assistants.
Claude
Paste the same question into a fresh chat. Check whether web search is available and enabled, then look for visible evidence that it ran. Claude’s web-search help explains its search indicators and citations.
Save the whole answer and any source links. If web tools are unavailable, note that limitation; don’t treat an answer without current search as equivalent to a searched local-business answer.
Gemini
Use the same prompt and explicit location. Note known location permissions, account context, and the mode shown. Google’s Gemini privacy guidance describes location information and personalization that can affect relevance.
Save the answer and links shown. If the answer uses a different location from the one you supplied, record the mismatch instead of quietly correcting the prompt for Gemini alone.
Perplexity
Start a new thread, paste the same question, and record the search mode and model if visible. Save the answer and source links. Perplexity’s explanation of how it works describes its numbered source citations.
Pay particular attention to citation versus recommendation. A cited page may support advice about a service without its business being offered as a provider. Read the sentence and the recommendation list, not just the source panel.
Save comparable evidence
Create one record per answer in a document or spreadsheet. Save:
- The exact prompt, without editing it afterward.
- The engine, date, time, and time zone.
- The mode or model shown, search status, and known location settings.
- The full answer text, including caveats and qualifications.
- Every business named—not only yours.
- Your business’s appearance label and any incorrect facts.
- All visible source links and which claims they seem to support.
Copy the answer text and URLs while they are available. A screenshot can supplement the record, but a cropped image is not a substitute for the full answer. Keep evidence in your own files rather than making a public chat link just to save it.
Use this blank comparison grid for each exact question; these are fields to fill, not sample results:
| Engine | Appearance label | Incorrect facts? | Businesses named | Sources and settings record |
|---|---|---|---|---|
| ChatGPT | — | — | — | — |
| Claude | — | — | — | — |
| Gemini | — | — | — | — |
| Perplexity | — | — | — | — |
Attach or reference the full answers beneath the grid. Open important citations and check whether the page concerns the right business and supports the claim. A source link is something to inspect, not proof by itself.
If you prefer a first recorded observation, the free Honest Visibility scan records one unbranded customer search behind a private link, with no account required. It is a starting point, not a substitute for this four-engine comparison.
Compare results across engines
Agreement and disagreement both tell you something
If all four assistants omit your business for the same relevant question, that is a stronger reason to investigate than one miss. It still does not identify the cause, and the four answers are not four independent votes proving a universal result.
If only one includes you, keep that observation alongside the other three. Don’t declare success and stop—or call the other engines broken. Inspect differences in search use, geography, sources, and interpretation of the customer’s need.
If an answer gives only general advice and names no businesses, note that answer style. It is different from an answer that recommends alternatives but leaves you out.
Separate sources from recommendations
Your site being cited does not mean your business was recommended. A directory being cited does not mean your official site was used. A competitor being named does not prove its website caused the selection.
You may notice that another business clearly explains a service or location your site leaves vague. That is a useful clue to investigate, not a demonstrated reason the assistant chose it. Don’t turn these observations into an invented visibility score or stable rank.
Give incorrect facts priority
Wrong hours, a wrong city, an unsupported “closed” label, or a service you don’t offer can mislead a customer even when you appear.
Verify the fact against your actual business information. Correct your official site where needed, then important profiles or listings carrying the error. Save the original evidence and correction date. Re-test afterward with the same question and comparable settings, without assuming the correction will immediately change every answer.
What to do with what you found
If you appear accurately for questions that matter, keep the evidence and maintain your core facts. You don’t need to invent a marketing project.
For mixed results, finish your preselected questions before adding more. If a relevant customer need was missing, add one deliberate variation and use it across all four engines. A recorded multi-engine pass may help if the result will inform a real decision.
For absence or errors, use Why AI Doesn’t Recommend Your Local Business. Investigate access, clear service and location facts, and credible corroboration. Don’t respond by publishing piles of AI-written pages. Crawl access, structured data, and profile corrections do not guarantee recommendations.
If this raises budget anxiety, AI Visibility vs. SEO explains why useful search work still matters. Improve information customers need rather than abandoning it for a new label.
Manual multi-engine check versus a recorded scan
A manual check can be free, subject to the assistants’ access and usage limits. It is flexible, and you control the questions. You also have to preserve the evidence and avoid prompt drift yourself.
Honest Visibility runs its recorded checks through search-enabled provider APIs—automated connections to the providers—with consistent instructions. It does not reproduce your personal chat account or its location and personalization settings, so its answers may differ from your manual checks.
The free Honest Visibility scan records one unbranded AI customer search, with a private report link and no account required. Start there if one recorded observation would be useful. Review it before deciding whether you need anything else.
The optional complete analysis is $29 one-time, with no subscription. It includes 20 responses across ChatGPT, Claude, Gemini, and Perplexity: three customer searches across four assistants produce 12 responses, and two business-accuracy checks across four produce another eight. It also includes supported competitor website comparisons and one prioritized plan.
Those two accuracy checks deliberately supply your business name, or your name and website. They assess claims against checked website evidence, flagging conflicts and information that cannot be verified. They do not verify every business fact, and their responses do not count toward unbranded appearances.
Comparisons cover up to three qualified competitor websites when they can be confidently matched and reviewed; they are not proof of why an assistant selected a competitor. The complete-analysis details explain the scope.
That recorded pass may be useful when one engine or one answer isn’t enough for the decision you’re making. It remains a point-in-time analysis, not continuous monitoring or a promise of future recommendations.
A short checklist to keep
- Choose an unbranded customer question and a small prompt set.
- Ask each exact question across all four assistants.
- Record search use, known location context, and full answers.
- Separate recommended, mentioned, cited only, and not present; flag errors too.
- Investigate supported problems, then re-test deliberately.
Start with the smallest useful observation. The free scan is an option; the $29 complete analysis is there if you need a broader recorded check. Neither replaces the basic discipline: ask fairly, keep the evidence, and act on what it actually supports.