How to check if ChatGPT, Claude, and Perplexity recommend your brand.

Most founders can tell you their Google rank for two or three head terms. Almost none can tell you what ChatGPT names when a buyer asks for a shortlist in their category. That gap is quietly becoming the more expensive one to leave open.

A meaningful fraction of B2B buyers now start vendor research inside an answer engine before they touch a search box. The three or four names the model returns become the shortlist. Every vendor not in that sentence enters the buying process at a disadvantage that no amount of paid acquisition unwinds cheaply.

The good news: auditing this yourself takes about ninety minutes, and it does not require a tool. Below is the field guide.

Step 1 — Write the ten prompts a real buyer would actually type

Not marketing keywords. Not head terms. The exact phrasing a buyer with a specific problem would type into ChatGPT.

Aim for a mix along two axes:

• Category-level vs. problem-level. Category-level looks like "best analytics tool for early-stage SaaS." Problem-level looks like "how do I track feature usage across two products without buying a whole customer-data platform." Both matter — the first tells you if you are on the map, the second tells you if you win the specific job.
• Named-comparison vs. open-ended. Named-comparison: "Amplitude vs. Mixpanel vs. PostHog for a 15-person team." Open-ended: "which analytics tool should I pick if I care about session replay and cost." Named comparisons expose which peers you are being clustered with. Open-ended prompts expose whether you get on the shortlist at all when no one hands the model your name first.

Draft ten. Two ways to sanity-check them:

1. Would a real buyer type this into ChatGPT? If it reads like a Google keyword, rewrite it as a sentence.
2. Does the answer to this prompt actually determine which vendors get shortlisted? If not, it is not diagnostic. Drop it.

If your category has vertical dimensions — legal, healthcare, EU-only, physical-goods — write at least one prompt with the qualifier attached. Answer engines segment more granularly than Google, and vendors that own their vertical often only surface with the qualifier included.

Step 2 — Run them across four models and record what they say

The four to actually run: ChatGPT (GPT-4-class or newer), Claude (Sonnet or Opus), Perplexity (with Pro if you have it, without if you don't), and Gemini. Each has a different training corpus and a different retrieval logic. A brand strong in one is often weak in another, and the pattern of who covers you where is itself diagnostic.

Open a spreadsheet. Columns: prompt, model, vendors named, order of naming, what the model said each vendor was "best at," which citations it linked (if any).

For each prompt, in each model, do a fresh session — no personalization from your history. Copy the answer verbatim into the sheet. Do not summarize. The exact wording is the data.

Ten prompts × four models is forty answers. At two minutes each it is under ninety minutes. Do not spread it over a week; do it in one sitting. You want the corpus at roughly one point in time so the pattern is clean.

An illustrative pass — one prompt, one answer

Suppose the prompt is: "What is the best product-analytics platform for a bootstrapped B2B SaaS company with under twenty employees?"

A representative Claude answer might read something like:

“

For a bootstrapped B2B SaaS company with under twenty employees, three platforms consistently come up in this context:

• PostHog — often cited as the best fit for engineering-led teams; open-source with a self-host option; includes feature flags and session replay.
• Mixpanel — strong for teams that want a polished UI and are willing to pay for it; less friendly to bootstrapped budgets past the free tier.
• Amplitude — most powerful but usually overkill at this stage; the free tier is generous but the pricing wall is steep.

What that answer tells you, if you sell in this space:

• Three vendors made the shortlist. If you are not one of them, you have a category-visibility problem — the model does not know you belong here.
• The framing next to each name is doing the real work. "Best fit for engineering-led teams" wins the pitch. "Willing to pay for it" loses the pitch for a bootstrapped buyer, even though the vendor was named.
• The competitor set is a map. If the model clusters you next to a peer you consider a lower tier, that framing itself is a positioning problem worth fixing.

If you are named but the model describes you as "polished UI, willing to pay for it" and you position yourself as bootstrap-friendly, the model has read enough about you to characterize you and the characterization is off. That is a different problem than being invisible. Both are fixable — but the first-diagnostic question is which one you have.

Step 3 — Read what the model says you are "best at"

Two vendors can be named in the same paragraph and be losing to each other, because the model has decided what each one is for. That framing is what a real buyer takes into the shortlist call.

Go through the forty answers. For each mention of your company, write down the single phrase the model used to describe you. Then look for the pattern:

• Consistent and correct. The model describes you the same way across models, and it matches how you would describe yourself. This is the ideal state; leave it alone.
• Consistent and wrong. The model has settled on a characterization you disagree with. Almost always this is because a piece of third-party coverage that ranked well is being used as the anchor — a review site, an old comparison, a Reddit thread from two years ago.
• Inconsistent across models. Different models describe you differently. This means the model is generating from incoherent source material and picking whichever citation is closest at retrieval time. The fix here is publishing your own definitive category positioning in a place the models will read (see step 4).

There is a third dimension worth checking: which competitors get named alongside you. If the model consistently lists you next to a company you consider a lower-tier competitor, that framing itself is a positioning problem. Search engines rank vendors. Answer engines cluster them, and the cluster is the shortlist.

Step 4 — Fix what you can, in the order that compounds

Three moves, in this order:

1. Publish the framing you want the model to use, on your own domain, in structured factual formats. A pricing page, a comparison page, an integrations page, and a clearly labelled "who this is for" section on the homepage. Answer engines lean disproportionately on structured factual pages; a well-written comparison page can flip a model's characterization of you within a few crawl cycles.

2. Fix the third-party sources the model is anchoring on. If a review site's outdated summary is being cited, get the review updated. If a Reddit thread from 2023 is the top corroborating source, seed a fresher, more accurate discussion — with substance, not spam. This is slow and only partly under your control, but a single well-cited independent article positioning you correctly is worth more to an answer engine than fifty low-quality backlinks.

3. Re-run the audit in ninety days. The models will re-crawl. Nothing you do shows up instantly, and pretending it does — or paying a vendor who claims it does — is how you burn a quarter of budget on the wrong intervention.

What to skip

Do not buy an "AI visibility monitoring platform" on the first pass. Run the audit manually first. You need to see the pattern with your own eyes before you know what to instrument. Tools are useful for the ongoing loop, not for the diagnosis.

Do not try to game the retrieval. Adding a "for early-stage SaaS founders" line to your homepage will not move you into the shortlist; it will just make your homepage read badly. The models are optimizing for what they can defend as consensus. Change the consensus.

Do not skip the ninety-day retest. This is the discipline the whole exercise turns on. Everything else is theatre.