AI visibility: can customers find you and understand what you do?

Start with one buying question and follow the answer to its source. Check whether it describes the right business, gets the facts right and gives the reader a useful next step. That tells you more about what to fix than a screenshot showing your name in an AI answer.

Answer engine optimization, or AEO, is a common label for work intended to improve how a business appears in AI answers. The work can involve public information, website accessibility, supporting evidence and the customer experience after a referral. It does not buy control of the answer.

This guide is for a founder deciding whether there is a real visibility problem and what kind of help it needs. You can do the first check yourself.

Separate four different questions

Being mentioned, being described accurately, being cited and being chosen are different things. Keep them separate in your notes.

These are diagnostic questions we use to choose work. They are not disclosed ranking factors or a formula for winning citations. A company might pass the first three checks and still send a prospective customer to a confusing service page.

Make your first observation reproducible

Choose questions people ask while deciding whether to buy. Use your own conversations and permitted customer research. A question written to make your company the obvious answer is a poor substitute for that work.

Separate branded questions, such as what a named company offers, from unbranded questions about a need or category. Record the exact wording, date, product, search setting, location assumption and model where the interface identifies it. Save the answer and its cited URLs. Check the sources yourself.

Repeat comparable observations on another date. If you change the prompt or test location, record that as a different test. A single absence is a finding about that answer, not proof that an entire platform cannot find you. A small sample is not a market-wide visibility score.

Use four labels for accuracy: correct, incorrect, incomplete and unknown. Unknown matters. You may be unable to identify where an unsupported sentence came from.

Trace the error before changing the page

Imagine a fictional studio called Alder. It designs retail interiors but does not manage construction. An assistant describes it as a design-and-build contractor and links to an old directory profile.

The owner checks the current website, the directory and the answer. The site is accurate; the directory still describes an abandoned service. The useful next action is to request a correction to that profile and make the delivery boundary easy to find on the service page. Publishing ten more articles about retail design would not address this particular error.

Record the correction date. Repeat the original question later and inspect the new sources. If the answer changes, report what you observed. Do not conclude that your edit alone caused the change: the assistant may have used different sources or other information may have changed.

If every checked source is accurate, leave room for an answer-generation error. You do not need to distort the website to match a faulty summary.

Access matters, but access is not a recommendation

Ask whoever maintains the site to check the relevant public pages and crawler controls. Do not broadly remove protections or expose private documents to chase visibility.

Google's current guide connects AI visibility to its search fundamentals. Pages need to be indexed and eligible for snippets, and the site must be included in Search generative AI features through Search Console's control. Special AI markup is not required; eligibility does not guarantee inclusion. Ask the site's owner to check the actual inclusion setting before assuming a content problem. Google's current guide, Search Console inclusion control

OpenAI distinguishes discovery through OAI-SearchBot from GPTBot training controls. A training preference should not be confused with the choice to make public pages available to search. OpenAI publisher guidance

Perplexity documents separate behavior for its search crawler and user-request fetcher. Check its current instructions rather than assuming one bot rule covers every type of access. Perplexity crawler documentation

Keep useful answers; retire unsupported shortcuts

A clear question-and-answer section can help a person decide. That is a good reason to write one. It is not a guarantee that a platform will cite it.

Google's documentation records the retirement of FAQ rich results from May 7, 2026. It also says `llms.txt` does not improve or reduce Google visibility or rankings. Keeping such a file for another documented use is a separate decision. Google documentation updates

Apply relevant structured data accurately and keep it consistent with the visible page. Do not add claims, reviews or services that a reader cannot verify there. Prioritize wrong facts, inaccessible pages and missing evidence before optional files whose benefit you cannot demonstrate.

Measure the customer journey as well as the answer

Your observation log can track correct descriptions and citations for a defined question set. Website analytics may identify some referred visits. An enquiry review can tell you whether the people contacting you are a fit. These measurements answer different questions.

Google now documents a dedicated Generative AI performance report for impressions from AI Overviews and AI Mode. Inspect what is available for your property and whether it contains enough data. That report does not establish AI-specific clicks, conversions or revenue. Overall Web reporting also includes AI-feature traffic; do not relabel its totals as AI Overview results. Generative AI report documentation, overall Web reporting guidance

For a referral, inspect the landing page and the action it offers. A visit with no enquiry can mean many things: early research, poor fit, a confusing offer, a broken form or a decision made elsewhere. A source tag cannot tell you everything that influenced the buyer.

Define the baseline and review window before making the change. With little traffic or a long buying cycle, the conclusion may remain inconclusive. It is still useful to correct a factual error; you do not need to invent a revenue lift to justify accurate information.

Choose one next step

Write down the question, the observed problem, the supporting source, the smallest justified correction and who can make it. Include one competing explanation and the date you will check again. If no problem is established, record that and leave the page alone.

If the answer is inaccurate, investigate the facts. If the source is weak, improve the evidence. If the customer cannot act, examine the journey from a visit to a qualified enquiry. The website improvement guide shows how to carry a justified change through review and follow-up.

Bring one answer and the page behind it to Brand Therapy if you would like a second reading. We can help identify which part deserves attention.

Platform guidance checked September 15, 2026. The Alder example is fictional; the diagnostic and sampling method are 8GNC editorial recommendations, not platform requirements.