AEO for agencies: how to run it and report it to clients
In short
Agree each client's tracked questions up front, measure them weekly on each assistant, and work a short list of changes a month in the order that moves answers. Report three numbers with their change and noise band, the work shipped, and a verdict on each change against questions it didn't touch. That last part is what keeps the retainer: it shows the client what the work did, not just what the market did.
Keats, from PressRoom. Updated Read as markdown
Clients are asking their agencies about ChatGPT. Some ask because their traffic dipped, some because a competitor's CEO bragged about it, some because they asked an assistant about their own company and didn't like the answer. Agencies that can answer with a number, a plan and proof keep the conversation; the rest get asked why they didn't see it coming.
What to sell
AEO fits into the retainers agencies already run. A useful way to frame it for a client:
- SEO agencies: the same pages, a new reader. Assistants search before they answer, so rankings feed citations; AEO adds the measurement and the third-party work.
- Content agencies: one page per buyer question, written answer first, and a way to show which pages got cited.
- PR agencies: the roundups, reviews and publications that assistants cite for buying questions are earned media. AEO tells you which placements change the answer.
Sell outcomes (named more often on the questions that decide sales) rather than activities (prompts tracked, pages optimized). The activities follow.
Setting up a client
- Agree the questions. Between twenty and thirty for each product line, phrased as buyers phrase them, mostly buying questions, with the client's brand left out of them. Get sign-off: the questions are the yardstick, and changing them later resets the baseline.
- Agree the rivals. Two to five competitors the client actually loses deals to.
- Agree the assistants. Usually ChatGPT, Gemini, Perplexity and both of Google's AI features. Some B2B clients add Copilot.
- Take a baseline. At least two or three weeks of answers before the first change, so you have something to measure against.
- Fix the plumbing. Crawler access in robots.txt and the CDN, and any wrong facts already being repeated. These are quick wins clients notice.
The monthly cadence
| When | What |
|---|---|
| Weekly | Read the numbers; note moves outside the noise band; catch wrong facts |
| Weekly | Ship the next change from the list, in order: fix, update, write, place |
| Monthly | Client report, with verdicts on changes that reached day 21 or day 35 |
| Quarterly | Review the question set and rivals with the client |
The order of work is the same for every client: correct what assistants get wrong, update pages they already read, write the pages they're missing, then get onto the third-party pages they cite. Our AEO strategy guide sets out the first 90 days.
The client report
A report that keeps a retainer fits on one page:
- The month in one sentence. "Prepco is named in 46% of buying answers, up 9 points since September and outside the noise band; the leader is at 61%."
- Three numbers, each with its change and noise band: named, named first, cited. If you show a single AI visibility score, show the three under it.
- Beside the rivals, on the same questions.
- What shipped, with dates and the questions each change targeted.
- Verdicts. For each change that's had three to five weeks: how far its target questions moved against the questions it didn't touch. Our guide to proving AEO results covers the method.
- Next month's work, in order.
Leave out screenshots of single answers unless they show a wrong fact. One answer is an anecdote, and clients who see a different answer the next day stop trusting the report.
Honesty is the retention strategy
Some changes won't move anything. Report them. A verdict that says "no measurable effect; we've revised the page and will read it again" is more valuable to a client than a vague claim that everything helped, because it means the verdicts that say "worked" can be believed. Flat inside the noise band is a result, not a failure to hide.
Likewise, don't promise rankings in ChatGPT. Nobody controls the answers. What an agency can promise is the method: the right questions, honest measurement, work in the order that moves answers, and proof of what moved.
Add the data clients already have
Search Console's generative AI report, rolled out to all sites in 2026, shows impressions from Google's AI features by page. GA4 can group visits from assistants into one channel. Both belong in the appendix of the report: they show reach and traffic, while the tracked questions show whether the client is named. Our guide to measuring AI visibility explains how the pieces fit.
Running many clients
The work per client is mostly reading, deciding and writing, which scales badly with headcount and well with an agent. With Keats, each client gets its own workspace on the Agency plan, with shareable report links and PDFs, and an AI agent can run each client's weekly cycle through the API, with a strategist approving what ships. More on that on our page for agencies.
Sources
- Keats: llms.txt, the operations and workflows for agents, read November 13, 2026
- Google Search Central blog: Introducing Search generative AI performance reports in Search Console, read November 13, 2026