What AI visibility means for agencies and their clients

For agencies, AI visibility is a new line of reporting and a new line of advice: whether ChatGPT, Claude, Gemini and Perplexity recommend the client, cite their site, and describe them correctly. It rewards agencies that measure honestly, explain uncertainty and focus on the sources assistants read, and it punishes those who promise rankings.

Clients are starting to ask the question. The agencies that answer it well have a method, not a screenshot.

Why are clients asking about it?

Because they, and their customers, use assistants. A buyer who asks an assistant for a recommendation may never see a search results page. Clients notice when a competitor is named in an answer and they are not, or when an assistant describes them wrongly. For an overview, see what AI visibility is and why it now decides who gets chosen.

The agency's value is in turning an anxious question ("are we in ChatGPT?") into a clear picture and a plan.

What is different from traditional SEO reporting?

Several things, and being direct about them builds trust.

• No rank #1. Answers are written fresh, and the same question can give different brands from run to run. There is no stable position to report. See why there is no real "rank #1" in AI answers.
• Variation is normal. One answer is a sample. Reporting requires repeated runs and stated uncertainty.
• Four assistants, not one. They search differently and recommend differently, so results should be shown per assistant.
• Third-party pages matter a lot. Much of the work happens off the client's site.
• Roles matter. A mention is not a recommendation. Mentioned, recommended or cited explains the difference.

Search skills carry over: clear content, technical access, authority. The measurement is what changes.

What should an agency measure for a client?

Keep it to a small set that means something. AI visibility metrics that actually mean something covers it in detail. In short:

1. A fixed question set, written from what the client's buyers ask, with no brand names. How to find those questions.
2. Recommendation rate per assistant: how often answers positively recommend the client.
3. Citation: how often the client's own site is cited, and which pages.
4. Presence on the pages assistants cite: whether the roundups and reviews list the client, accurately.
5. Competitor comparison: the same measures for two or three rivals, so the picture is relative.
6. Factual accuracy: errors found in answers about the client.

Run it on a regular schedule, saving every answer. Tracking over time gives a routine.

How should you report it?

Plainly, with the evidence attached.

• Start with the answer to the client's question. The shape to aim for is "In X of your N tracked questions, at least one assistant recommended you", with real figures you can trace.
• Show the per-assistant grid, not just an average.
• State the uncertainty. If a change is within normal variation, say "no clear change".
• Link the numbers to answers. A client should be able to see what the assistant actually said.
• Separate what you did from what happened. Be honest that influence is indirect and slow.

Reports built on a single run, or on a vague composite score, invite the question "what is this made of?" and rarely survive it.

What do you actually do for the client?

The work is concrete, and mostly familiar.

• Fix the basics: a clear homepage, honest pricing, crawl access. See why your homepage confuses AI assistants and AI crawlers and robots.txt.
• Correct the record: find wrong or outdated claims and trace them. See how to fix outdated information.
• Earn presence on cited pages: roundups and reviews, with an honest pitch.
• Create content that answers real questions, including fair comparisons.
• Support real customer voices through reviews and communities, never fake ones.

What should an agency never promise?

• A ranking or a position. Nothing like that exists in AI answers.
• A timeline. Discovery and recommendation take uncertain time.
• Guaranteed inclusion. No one outside an assistant controls what it says.
• Results drawn from a single run. A screenshot of one lucky answer is an anecdote.
• A magic score whose inputs you cannot explain.

Promise a method, evidence, honest reporting and sustained effort. Those are things you control.

Where does Citewake fit?

How do you price and package the work?

Be clear about what the client buys: a measurement routine, a set of fixes and a report, not a result. A typical package has a baseline, a regular check against a fixed question set, a short action list and a review after a few cycles. Keep scope tied to things you control. That makes the engagement easy to explain and hard to be disappointed by.

How do you start?

Pick one client and 20 to 30 buyer questions. Check them across the four assistants, save every answer and write up what you see. That first pass usually produces enough findings, such as wrong facts, missing listings and unread pages, to show the value of doing it properly.

Citewake builds an index of AI answers and publishes how it measures them, so that every number can be traced to the words an assistant used. You can read the approach in the methodology, and try the Citewake checker for a first look at a client's site. Agencies can use the same standards in their own reporting: fixed questions, kept answers, stated uncertainty.