How to — find out what AI answer engines say about you

You cannot see what ChatGPT, Gemini or Perplexity say about your company until you go and ask them the way a customer would. This is a two-hour audit that produces a number you can defend, instead of a vendor dashboard score nobody can reproduce.

It also tells you the more useful thing: which questions you are losing, and to whom.

1. Write down the questions, in the customer's words

Not keywords. Questions. Fifteen to twenty of them, phrased the way a person types when they are deciding whether to buy from you: is this company any good, how much does it cost, who does this in Singapore, this tool versus that one, what are the alternatives, is it safe for a small team.

If you have support logs or a sales inbox, the phrasing is already there — lift it. The single strongest page-level predictor of being cited is how closely a page's wording matches the questions its audience actually asks, according to a study of roughly two million AI citations across four commercial engines published on arXiv in September 2026, which found that alignment beat every other page feature tested. You cannot match wording you never wrote down.

Put the list in a spreadsheet: one row per question, one column per engine.

2. Ask in a clean session, and record the answer verbatim

Log out, use a browser profile you do not normally use, and start a new chat for each question — memory and history change what comes back. Paste each answer into the sheet, unedited. Do not summarise it; the summarising is where people accidentally write down what they expected.

For each answer record four things: did you appear at all, were you linked or only described, was every fact about you correct, and which sources were cited instead. That last column does more work than the other three combined. It is your competitor list for that question, written by the engine.

3. Ask each question several times, on different days

One answer is not a measurement. Work published in March 2026 and revised in August sampled three engines — Perplexity, OpenAI's search product and Gemini — daily for nine days and again at ten-minute intervals, and found that identical queries returned different answers and different citations, that citation distributions were heavy-tailed, and that rankings were unstable across runs, the author reports. The paper's blunt conclusion is that single-run visibility numbers look precise and are not.

Practically: five samples per question, spread across a week and across times of day, is the smallest set that lets you say "we appear in three of five" rather than "we appeared once". If a competitor shows up in one sample out of five, you have learned nothing about them except that they were lucky that morning.

4. Read the pages the engine cited instead of you

Take the two or three pages that keep winning your lost questions and read them the way the engine does: is the answer to the question in one plain sentence near the top, does that sentence carry a number or a named source, and is the rest of the page about exactly that question rather than the general category?

This is where a lot of GEO advice falls apart. The same two-million-citation study tested more than sixty page features and found that the standard answer-engine checklist — FAQ blocks, structured data, fast page loads — looked positive in pooled data and then collapsed to nothing, or reversed, once the researchers controlled for which domain the page lived on. Domain-level authority outweighed the strongest page-level feature by a factor of six. The engine is largely deciding whose site to trust before it reads your page, and the two levers you actually control are the wording of your answer and who else on the web is willing to cite you.

5. Check whether the crawler can reach the page at all

An embarrassing share of "we don't show up" cases are retrieval failures, not writing failures. Each engine publishes the identity it crawls under — OpenAI lists its crawlers on its bots documentation page, and Google lists its common crawlers, including the one used for AI features, on its crawler reference. Check three things against the page you care about: your robots rules do not block those agents, the content is not behind a paywall or a login, and the text is present in the HTML rather than assembled after the page loads by a script.

Then open your server logs and look for those crawler names hitting that page in the last thirty days. A hit is the cheapest evidence you will ever get that the page is in the index at all. No hit means fix access before you rewrite a single sentence.

6. Fix one page, then rerun the same questions in two weeks

Change one page — usually the one losing the most questions — and leave the prompt list alone, because rewriting the questions resets the comparison. When you rerun, report three numbers separately: how many questions name you, how many link to you, and how many sessions arrive from those engines in analytics.

Keep them separate because they do not move together. A study of Google's AI Overviews compared English Wikipedia's search referrals with the same articles in German and French as the feature rolled out country by country, and found default availability cut English search traffic by 5.45%, per the paper. Being quoted and being visited are different outcomes, and a citation is closer to branding than to traffic.

Don't do this: buy a GEO score and make it your KPI

Composite visibility scores are built on single-run sampling with unpublished weighting, which is exactly the measurement the March paper shows to be unstable. You cannot tell which of your changes moved a score you did not compute, and neither can the vendor, so you end up optimising a number that mostly reflects how many prompts their panel happens to contain.

The audit above is slower by an afternoon and better in every other way: you can hand anyone the prompt list and the raw answers and they can check your work. Read our explainer on what generative engine optimization actually is before you accept any vendor's framing of the problem, and if your answer engines are citing thin pages that look machine-written at scale, why AI slop stopped working is the other half of the story.

How you'll know it worked

The same twenty questions, run five times each, two weeks later: you appear in more of them, you are linked rather than silently absorbed, and the facts the engine states about your pricing and your products are correct. Your analytics shows sessions from those engines landing on the page you fixed.

If citations climb and clicks do not, the audit worked and the market moved. Say so out loud before someone reports it as a failure.

Have you run this on your own company? Tell us the worst thing an answer engine got wrong about you. Write to us via our contact page.

How to — stop an agent before it burns your budget

An agent that runs unattended fails in a way a chat window never does: it fails expensively. The job here is to make the worst case of any run a number you know before you press go — and to have the run stop itself the moment it stops making progress. Most teams discover this the hard way. A task that should take twelve steps takes four hundred, half of them re-asking the same question in slightly different words, and the invoice arrives two days after the incident.

How to — check a model's license before you ship on it

You want to know, before you ship, whether the weights you downloaded oblige you to ask anyone's permission — and whether that answer survives your fine-tune, your hosting plan, and next quarter's version bump. It's a reading job with a fixed set of questions, and it takes about fifteen minutes. The reason to do it now rather than when legal asks: the terms moved this quarter.

How to — review AI-written code before you merge it

You want to merge AI-written code without reading every line of it — and still be able to say, out loud, what it does and what happens when it breaks. That is a review job, and it is mostly about reading the right parts in the right order. Cockroach Labs ran its codebase like a teaching hospital for five months and reports merging well over a million lines of code with seven total reverts — a run where no agent's plan moved without a human-signed review first (we covered Coding agents as a…

How to — estimate what an AI feature will cost before you ship it

You want one number you can defend in a planning meeting: what each completed job on your feature costs, and what a month of it costs at launch volume. This is a measurement job, not a guess — a morning with your real prompts beats any pricing page. 1. Run 20 real requests and keep the usage numbers. Not demo prompts — your actual system prompt, tool definitions, the documents you stuff into context, and a typical user message.

How to — check an AI-generated citation before you rely on it

An AI-drafted citation can point at a real document and still misdescribe it — that is the failure mode a link check misses. Verification means opening the source and reading the sentence your citation is attached to. This is a job you can finish in an afternoon for a normal reference list, and it is the difference between using AI as a research assistant and signing your name to something you never read. 1.

How to — keep your AI feature alive when a model is retired

Every model you call has a shut-down date, and most teams only learn it from a failed request in production. The dates are published months ahead. The breakage happens anyway. Treat retirement as a scheduled maintenance window rather than an emergency: know what you depend on, know your provider's clock, and have the replacement tested before the deadline arrives. 1.

How to — cut your LLM bill without switching models

Your token spend keeps climbing, and the obvious fix — swap in a cheaper model — is the one you should try last. In a live AI feature, most of the waste isn't the price per token. It's paying full price for tokens the provider has already seen, and generating output tokens nobody asked for. Work the list in this order. The first three moves usually move the number more than a model swap would, and none of them cost you quality. 1.

How to — build a small eval set for your own AI feature

You changed a prompt, and now you can't tell whether the feature got better or you just remembered the good answers. An eval set is how you stop guessing: a fixed batch of real cases with a written pass mark, run again every time you touch anything. You can build a usable one in an afternoon. There is nothing ceremonial about this. A good eval for a product team is 30 to 100 real cases, each with a short statement of what a pass looks like, kept in a file that gets re-run.