AI visibility for agencies: tracking and reporting ChatGPT recommendations for clients

Clients have started asking a new question: "Does ChatGPT recommend us?" For an SEO, content or growth agency, that's an opening. AI visibility work builds on what you already do (search, content, digital PR, reviews), and it can be measured and reported. This guide covers how to scope it, measure it honestly, and report it so clients understand what they're paying for.

What "AI visibility" means for a client

For a SaaS client, AI visibility comes down to a few measurable questions:

1. When buyers ask AI assistants for recommendations in the client's category, is the client named?
2. Where in the answer: first, or an afterthought?
3. Which competitors are named instead?
4. Which sources do the answers cite, and is the client on them?
5. Is the client described accurately: name, category, pricing?

Everything you do for the client should move one of those, and your reporting should show them over time.

Why it fits agency work

The levers that affect AI recommendations are mostly familiar agency services:

• Technical access and indexing. OpenAI says sites that opt out of its OAI-SearchBot crawler won't appear in ChatGPT search answers (apart from navigational links). Google says AI Overviews and AI Mode use the same Search fundamentals: pages must be indexed and eligible for a snippet. Perplexity runs PerplexityBot for its search results. Crawl audits now include these bots.
• Digital PR and listicle outreach. Search-based answers lean on third-party roundups and comparison pages. In our 26 October 2026 run of 100 ChatGPT answers across 20 SaaS categories, techradar.com alone was cited in 32 answers across 13 categories.
• Content. Comparison, alternatives and "best for [audience]" pages answer the questions buyers put to AI assistants.
• Reviews and directories. Profiles on review sites and directories describe the client in category language.
• Entity consistency. The same name, category and positioning everywhere, with matching structured data.

What's new is the measurement, and that's where agencies often go wrong.

Measuring it honestly

AI answers vary. In our 26 October data, 151 of 264 product-and-category pairs (57%) appeared in only one of five answers, and only 9 products appeared in all five. A screenshot of one chat is not a result, and a single "we're not in ChatGPT" panic isn't a crisis.

A defensible method:

1. Agree a fixed question set with the client. 15 to 25 questions in buyers' words: best-of, shortlist, alternatives to [competitor], [client] vs [competitor], "affordable [category] with [feature]", "best [category] for [audience]".
2. Agree the competitor set. Three to five direct competitors, not the whole market.
3. Run in clean conditions. Fresh sessions without memory or prior context, with web search, on a fixed schedule.
4. Record the same fields every time: named or not, position, competitors named, cited sources, accuracy of description.
5. Report trends over four to eight weeks, not week-to-week movements.

Include narrow questions on purpose. In our data, "Recommend an affordable [X] with good support" put a product outside the category's top 3 in first place in 16 of 20 categories. For smaller clients, narrow questions are often where the first wins show up, and that's a result worth reporting.

The baseline audit (week 1)

Start every engagement with a baseline. A good audit has five parts:

1. Visibility snapshot: share of answers, average position and first-place count for the client and each competitor across the question set.
2. Source map: every cited domain, grouped by type (editorial roundup, review platform, directory, forum, vendor site), with a note on whether each mentions the client, competitors, or both.
3. Accuracy check: ask "What is [client]? Who is it for? What does it cost?" and record errors such as old names, wrong pricing or the wrong category.
4. Technical check: robots.txt and CDN or WAF rules for OAI-SearchBot, PerplexityBot and Googlebot; Google-Extended settings; indexing in Bing and Google; whether key pages render as text.
5. Content gaps: comparison, alternatives and use-case pages competitors have and the client doesn't.

The source map usually becomes the work plan. Sources that cite competitors but not the client are the outreach list.

The monthly work plan

A typical retainer splits into four streams:

• Outreach: pitch the recurring cited roundups and comparison sites, with a factual product summary and test access. Track placements won.
• Content: one or two comparison, alternatives or use-case pages a month, fair to competitors, with direct answers and visible facts.
• Profiles and reviews: complete and correct review and directory profiles; set up an honest review request flow with the client's customer team.
• Technical and entity: fix crawler access, structured data, sitemaps and naming consistency.

Reporting that clients understand

Keep the monthly report to one page per client:

• Share of answers: "Named in 6 of 20 questions (baseline: 3 of 20)."
• Position: average position when named, plus how many answers named the client first.
• Competitor comparison: the same numbers for each competitor.
• Source changes: new cited sources that mention the client; sources still missing.
• Work done: placements, pages published, fixes shipped, each with a date.
• Caveats: a line noting that answers vary by phrasing, time and model version, and the numbers are a consistent sample, not every user's experience.

Show position next to share. In our product analytics data, PostHog was named in 3 of 5 answers, behind Mixpanel and Amplitude at 5 of 5, but it was named first every time it appeared. In email marketing, Brevo was named in 4 of 5 answers but never first. Clients should see both numbers.

Setting expectations

Be clear with clients about what AI visibility work can and can't promise:

• No guarantees of placement. No one controls what ChatGPT, Perplexity or Gemini say, and behavior changes with model updates.
• Lag. Third-party pages need to be published, indexed and then retrieved. Expect results to take weeks, not days.
• Sampling. Report on a fixed question set. Don't extrapolate to "ChatGPT always recommends us".
• Compliance. No fake reviews, no undisclosed paid placements presented as editorial, no hidden text. They break platform rules and put the client at risk.

Tooling

You can run all of this by hand with a spreadsheet: it costs time, not money, and the hard part is consistency. Tools automate the asking and the recording.

ShortlistWatch's Agency plan tracks 5 products for $79 a month: each product gets 25 buyer questions re-asked every week, with share of voice, cited sources, trends and a weekly "what to fix" email. The free check (5 questions, no account) is a quick way to produce a first snapshot for a prospect call. Today ShortlistWatch measures ChatGPT; for Perplexity and Gemini, use the same method by hand.

A simple engagement outline

1. Week 1: baseline audit and agreed question and competitor sets.
2. Weeks 2–4: technical fixes, profile cleanup, first comparison page, first outreach wave.
3. Monthly: two to four outreach targets, one or two pages, profile upkeep, one-page report.
4. Quarterly: refresh the question set if the client's positioning or market has changed, and re-run the full source map.

The short version

• Measure share of answers, position, competitors, sources and accuracy, on a fixed question set, repeatedly.
• Use the source map as the work plan.
• Do the familiar work: access, outreach, content, reviews, consistency.
• Report trends with honest caveats, and promise effort and measurement, not placement.

For engine-specific detail to share with clients, see our guides for ChatGPT, Perplexity and Gemini, and our walkthrough on checking whether ChatGPT recommends a product.