Navigating AI Search Visibility in 2026: A Guide for Agencies
Most agencies now get some version of the same client question: "Are we showing up in ChatGPT?"
The honest answer for most clients is "sometimes, on some engines, for some questions," and that's exactly why it needs a method. AI answers vary by engine, by phrasing and from one run to the next. A screenshot of one ChatGPT answer proves nothing either way.
This guide is the practical side: how to audit a client's AI search visibility, how to turn the findings into a diagnosis, and how to run the first 90 days of work. For the business side, packaging and pricing this as a service, see our companion post, AI Search Visibility: The Agency Frontier in 2026.
The prompt set is the foundation of the whole engagement. Everything you measure later depends on it, so spend real time here.
Start from buyer intent, not keywords. Keywords are short. Prompts are the questions buyers actually ask an assistant, usually with context. "Accounting software" becomes "what's the best accounting software for a small construction company that does job costing?"
Cover the funnel. A useful set mixes:
- Category questions: "What are the best options for X?" This is where shortlists form.
- Problem questions: "How do I fix Y?" These are where brands earn authority by being cited as the explainer.
- Comparison questions: "X vs Y for Z." These are close to purchase and often decided by third-party pages.
- Local or situational questions: "Who does X near Denver?" or "X for a team of 10." Specifics change the answer a lot.
- Brand questions: "Is [client] any good?" or "What does [client] cost?" These show whether engines describe the client accurately.
Size it sensibly. For a typical client, 25–100 prompts is enough to see patterns without drowning in data. Start with around 40 and grow it as you learn which areas matter.
Write them like real people do. Pull phrasing from sales call notes, support tickets, Search Console queries, Reddit threads and "People also ask" boxes. Avoid stuffing the client's brand name or marketing language into prompts. Buyers don't talk that way, and it will flatter the results.
Freeze the set. Once agreed, version it. You can add prompts later, but don't quietly edit existing ones, or month-over-month comparisons stop meaning anything.
Run each prompt on each engine your client's buyers use. For most clients that means ChatGPT, Google AI Overviews, Perplexity and Gemini at minimum, plus Claude, Copilot or Grok depending on the audience.
Because answers vary between runs, a single pass isn't a measurement. Run each prompt several times per engine, or across several days, and treat the result as a rate: "mentioned in 6 of 10 runs" is a finding; "mentioned once" is an anecdote.
For each answer, record:
- Mentioned brands, in order of appearance.
- Cited URLs and which domains they belong to.
- Whether the client was mentioned, and in what position.
- Whether the client's own pages were cited.
- How the client was described, including any wrong facts (old pricing, discontinued products, wrong location).
Doing this by hand for 40 prompts × 4 engines × several runs gets tedious fast, which is why most agencies use a tool for the collection and save the human time for interpretation.
Turn the raw answers into a few numbers the client can follow over time.
- Mention rate: the share of answers that name the client. This is the headline metric.
- Citation rate: the share of answers that cite at least one of the client's pages.
- Share of voice: the client's mentions as a share of all brand mentions across the prompt set. This is the clearest way to show competitive position.
- Average position: where the client appears when mentioned. First and fourth aren't the same outcome.
- Accuracy issues: a count of answers that describe the client incorrectly.
Break each metric down by engine and by prompt group. A client can be strong on Perplexity and invisible on ChatGPT, or win problem questions while losing every category question. Those splits are where the diagnosis comes from.
Numbers tell you where the client is losing. The cited sources tell you why. For each prompt group where the client underperforms, look at what the engines are citing instead.
If competitors' own pages are cited: their pages answer the question more directly. Compare them with the client's equivalent page. Typical differences: the competitor states a price range, lists specific use cases, includes a comparison table, or answers in the first paragraph instead of the fifth.
If third-party pages are cited: review sites, "best of" roundups, directories and forum threads are deciding the answer. Note which ones list the client and which don't. That becomes your outreach and listings plan.
If nobody relevant is cited: the engine may be answering from training data. That's slower to influence and responds to broad, consistent presence across the web more than to any single page.
If the client is described wrongly: find the source of the error. Often it's an outdated page on the client's own site, an old press release or a stale directory listing. Fix it at the source; engines repeat what they find.
Also check the basics while you're here:
- Crawler access. Check robots.txt and any bot-blocking rules at the CDN for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended. Decide deliberately which to allow. Accidental blocks are common.
- Indexation and rendering. Pages that need JavaScript to show their main content may be invisible to some crawlers.
- Entity consistency. Name, address, product names and core facts should match across the site, Google Business Profile, major directories and social profiles.
- Structured data. Organization, Product, Article and FAQ markup, where it matches visible content, help engines connect facts to the right entity.
With a diagnosis in hand, the first quarter tends to fall into three phases.
Days 1–30: Foundations and flagship answers
- Fix crawler access, indexation and rendering problems on priority pages.
- Correct any inaccurate facts at their source.
- Pick the 5–10 prompts that matter most commercially. For each, make sure one page on the client's site is the best direct answer: lead with the answer, add specifics (prices, ranges, steps, comparisons), cite sources, and add the relevant schema.
Days 31–60: Authority and coverage
- Work the third-party list. Pitch inclusion in the roundups and comparison pages that engines keep citing. Complete and correct directory and review-site profiles.
- Expand content for prompt groups where competitors' pages win, focusing on specific situations rather than generic overviews.
- Add original evidence where you can: client data, survey results, benchmarks, case studies with numbers. Evidence-dense content gets cited more often.
Days 61–90: Measure, learn, standardize
- Re-run the full prompt set using the same method as the baseline.
- Report changes by engine and prompt group, and connect them to the work done.
- Document what moved and what didn't, and set the next quarter's priorities.
- Agree on a re-measurement rhythm: monthly for most clients, weekly in fast-moving categories.
Here's how this plays out for a hypothetical client: a 12-location physical therapy group.
Prompt set: 48 prompts, including "best physical therapy for a torn ACL in Charlotte," "do I need a referral for physical therapy in North Carolina," "physical therapy vs chiropractor for lower back pain" and "[client] reviews."
Baseline: Mentioned in 18% of answers overall: 31% on Perplexity, 22% on Google AI Overviews, 9% on ChatGPT and 10% on Gemini. Never cited for the referral question, which a state government page and a competitor's FAQ answer instead. Two engines list a location that closed last year.
Diagnosis:
- The client's location pages are thin and don't mention conditions treated or insurance accepted.
- A local "best PT clinics" article cited on three engines doesn't include the client.
- The closed location still has a live directory listing and a page on the client's site.
First 30 days: Remove or redirect the closed location's page and listing. Rebuild location pages with conditions, insurance and appointment details. Publish a clear, sourced page answering the referral question for North Carolina.
Day 90 re-measure: The mention rate has moved, the closed location has disappeared from answers on engines that refresh from the live web, and the referral page is cited on Perplexity. ChatGPT has moved less, which is a useful lesson to share with the client about how engines differ.
The numbers in your own audits will differ. The value is in the method: a fixed prompt set, repeated runs, scoring, and a diagnosis based on what engines actually cite.
How many prompts should an audit include? Usually 25–100. Fewer than 25 makes it hard to see patterns; more than 100 adds collection cost without adding much insight for most clients.
How often should we re-run the audit? Monthly for most clients. Answers drift as models and search indexes update, so a single audit goes stale within weeks.
Why do results differ between runs? AI engines generate answers probabilistically and may retrieve different sources each time. That's why you measure rates across several runs instead of trusting one answer.
Should clients allow AI crawlers? For most businesses that want to be recommended, yes, at least for search-oriented crawlers such as OAI-SearchBot and PerplexityBot. Training crawlers are a separate decision. Make it deliberately with the client instead of inheriting a template's defaults.
Can AI visibility be measured in Google Analytics? Partly. Referrals from ChatGPT, Perplexity and others show up when the engine passes a referrer, but much of the influence happens without a click. Use referral data alongside answer-level measurement, not instead of it.
The difference between a one-off audit and a service is repeatability. The same prompts, the same engines, the same scoring, every month. That's what lets you show a client real progress instead of screenshots.
Foxcite handles the collection side of this workflow. It runs your prompt set across ChatGPT, Claude, Perplexity, Gemini, Grok and Google AI Overviews, records mentions and citations, and shows how each client's visibility changes over time.
Sources
