Your service business may be missing from recommendations, described incorrectly, or attracting people who want something you do not sell. Those are different problems. Each needs different evidence before you commission work.
Start with one service you want to sell and one question a prospective customer would ask before contacting a supplier. For an established B2B firm, include a real buying constraint: the problem, type of business, market or delivery requirement. Keep questions that name your business separate from questions that ask for a provider without naming one.
Use this page to choose the first investigation. Use the AI visibility checklist and one-page worksheet to record it. If you have a confirmed issue and need implementation, the AI SEO plan explains the work and its boundaries.
Start with the symptom you can actually observe
| What you observe | What to check first | What it does not establish |
|---|---|---|
| The assistant describes us when named, but leaves us out of a provider question. | Save both questions and answers. Check whether the provider question matches a service, market and buyer we actually serve. | Brand recognition does not establish discovery among unfamiliar buyers. One omission does not identify a website fault. |
| It gives the wrong service, location or business fact. | Compare the exact claim with its cited source and the correct business record. Identify who controls the inaccurate source. | The website is not necessarily the source of the error. An assistant can misread a correct source. |
| Competitors appear and we do not. | Record actual answers and cited pages across comparable observations. Inspect whether those sources address the buyer’s specific constraint. | A competitor mention does not reveal an algorithm’s weighting or prove that copying their page will work. |
| A key service page seems absent from search. | Ask the web team to check access, indexability, canonical selection and available crawl records for that URL. | Opening the page in a browser does not confirm crawler access through the host or CDN. |
| Search visits arrive, but enquiries are a poor fit. | Compare disclosed search wording, landing pages and the enquiries you can actually attribute. Read the page as someone choosing a supplier. | More impressions do not tell you who the searchers are or whether they intend to buy. |
If all you have is a screenshot, start by recording its context. Do not label the cause yet.
Make the observation useful enough to investigate
Write down the buyer question, service and market, whether the business was named, platform, visible model or search mode, date, full visible answer and cited URLs. Note whether the conversation was fresh and include relevant location or account context. Mark unavailable context unknown. Repeat comparable questions later and retain unsuccessful attempts and omissions as well as mentions.
For example, “Which HR advisory firms support UK technology companies hiring their first people manager?” describes a buying task. “Tell me about our company” checks recognition of a supplied name. These are example questions, not reported test results.
Answers need not be identical between users or runs. OpenAI describes how location and relevant saved memory can affect ChatGPT Search. A useful record preserves those conditions rather than treating one answer as a universal ranking.
For an incorrect answer, save three things together: the assistant’s claim, the source it cites, and the business fact with its authoritative source. If a directory is wrong, request a correction there. If your page is wrong, correct the page. If the source is right and the summary is wrong, record that distinction instead of rewriting correct facts to match the summary.
Check four layers, without guessing how an engine scores them
We use Discoverable, Understandable, Citable and Routable to organize an investigation. These are practical review categories, not a published search-engine formula.
Discoverable: can the relevant source be reached?
Check the intended service URL, its crawl and indexing controls, internal links and any verified access failures. Ask for dated evidence from the web team when a check needs account or server access. Mark it unknown until then.
For ChatGPT Search, distinguish OAI-SearchBot from GPTBot, which concerns potential training, and ChatGPT-User, which handles certain user-triggered visits. Review robots rules and host/CDN access with the technical owner. OpenAI’s crawler documentation explains the separate controls. Permission to crawl does not guarantee selection.
Understandable: does the page describe the buying decision?
Read the opening and service details. Can a buyer identify what you deliver, whom it suits, the markets you actually serve and the next step? Compare that language with navigation, related articles and relevant business profiles.
If two pages appear to sell the same service, establish their intended jobs before merging or redirecting anything. They may serve different buyers. If an article attracts people looking for software while you sell implementation, clarify the distinction where they make that decision.
Citable: can someone verify the answer and the evidence?
Locate answers about fit, process, scope, cost factors and delivery dependencies. Then inspect the proof: what was measured, over which period, from which source, and with what limitations? A result from one client is not a forecast for another.
“Citable” describes how clearly a source supports a statement. It does not promise that a system will cite it. Google’s AI features guidance does not require special AI markup or an AI text file for inclusion. Follow its eligibility guidance and keep structured data consistent with visible content.
Routable: can the right person take the next step?
Follow the journey from the article to the service, relevant proof and contact path. Try it on a phone. Make sure a form confirms success and reaches the intended recipient before treating a lack of enquiries as a demand problem.
Review enquiry fit alongside traffic. Where attribution is missing, keep it unknown. Ask whether the person needs the service, fits the delivery market and is ready for the next conversation. The published case studies show the evidence and context behind our own work.
A documented example from our own site
On 4 October 2026 we reviewed the Google Web Search evidence for our AI visibility cost page. This is a partial diagnosis of search intent on one page. It does not test AI answers, crawler access or every review layer.
Observed baseline. For 5 September to 2 October 2026, Search Console reported 1,047 impressions and 2 clicks for the page. Disclosed query wording included “ai visibility agency cost for b2b companies” and “ai visibility agency pricing and packages”, alongside “ai visibility platform pricing” and “ai visibility tool pricing”.
Interpretation to test. Those examples suggest different buying tasks: commissioning a service and comparing software. They justify checking whether the page makes our offer clear. They do not identify the searchers, establish why they did or did not click, or prove that the page caused unsuitable enquiries. Relevance, result presentation and competing results also need consideration.
Change prepared. At the date of this review, we had prepared a revision that separates software cost from agency work, explains service fit and routes relevant readers to the AI SEO plan. It was held for review and had not been published. The service and cost pages form the starting point for subsequent work.
Outcome still unknown. This review contains no post-publication result, AI-answer test or attributed lead improvement. Before publication, we need a consistent enquiry record and an agreed definition of a qualified enquiry: a relevant service need, customer fit and delivery market. Record the actual publication date, then compare relevant landing-page searches and qualified enquiries against a refreshed baseline, using comparable periods and recording other changes.
Read the dated evidence record. It includes the page filter, reporting period, selected query examples and limitations. Search Console dates are Pacific Time, rare queries can be withheld, and these are Web Search totals, not a separate AI-feature measurement.
The useful decision was to investigate the fit between the search question and the offer. The impression count alone would not have supplied that decision.
What should you fix, and what should wait?
Choose the next action from the evidence you have:
- A verified access failure or wrong business fact: give the specific URL, evidence and required correction to its owner. Recheck after the change.
- A visible mismatch between the buyer’s task and the page: clarify that page’s role and the relevant next step. Keep a copy of the previous version so the change can be reviewed.
- A missing answer or unsupported claim: add a useful answer or source where it belongs. Do not create a separate page for every wording variation.
- Unknown technical or measurement status: request the missing check. Unknown is not a reason to buy a rebuild or start a publishing programme.
For Google, check the Generative AI performance report if it is available for your property, alongside ordinary Search performance. Google documents AI Overviews and AI Mode impressions there. An unavailable report does not establish zero visibility. Our own-site example above uses only the Web Search data we inspected.
Track the issue corrected, publication date, observed search changes and enquiry quality separately. A mention, a click and a qualified enquiry are different observations. None should silently stand in for another.