How to choose a GEO or AEO agency (and test it yourself first)
By Shaeel Iqbal · Published October 14, 2026 · Last updated October 14, 2026
The short answer
Choose an agency that can show real, dated proof that AI names its clients, explains what it will change in plain words, measures results by repeated checks instead of one screenshot, and promises no guaranteed placement. Before you sign, test it yourself: ask AI your buyers' questions, from your buyers' country, several times, and see whether the agency's own clients show up.
We're an agency, so we have a stake in this. Score us with the same sheet below. We have one published case study, which is thin, and the scorecard asks about proof for that reason.
What an agency can and can't do
An agency can find out how AI describes your business today, fix what you control (your pages, facts and structure), and work on the mentions other sites make of you. It cannot control what an AI model says. Google's own guide says indexing and serving aren't guaranteed,[1] and AI tools rarely give the same list of brands twice: in SparkToro's study of 2,961 runs, ChatGPT and Google's AI gave the same list less than 1 time in 100 (not peer reviewed; run with Gumshoe.ai, which the author disclosed).[2] The research that gave GEO its name found that adding citations, quotations and statistics raised visibility by up to about 40% on one benchmark, and that keyword stuffing did not help.[3] Treat all of this as direction, not a promise.
Eight questions to ask, and what a good answer sounds like
1.Ask: "Show me a dated before-and-after from ChatGPT or Perplexity for a client in a similar field."
Good answer: Screenshots with dates, the exact question asked, and more than one run.
Warning sign: Only traffic charts or a tool's visibility score.
2.Ask: "Which questions will you track for us, and who chooses them?"
Good answer: The questions your buyers actually ask, which you approve.
Warning sign: A keyword list.
3.Ask: "How many times do you ask each question, and how do you report it?"
Good answer: Repeated runs, reported as how often you're named.
Warning sign: One screenshot.
4.Ask: "What exactly will you change on our site and elsewhere?"
Good answer: Specific pages, facts, structured data and listings.
Warning sign: "AI optimization" with no detail.
5.Ask: "Where will the mentions of us come from, and do you ever pay for them?"
Good answer: Real reviews, directories, earned coverage.
Warning sign: Bought links or reviews. Google says seeking inauthentic mentions isn't as helpful as it might seem.[1]
6.Ask: "Who does the work, and who will I talk to?"
Good answer: Named people.
Warning sign: Vague "our team".
7.Ask: "What happens if nothing moves, and what are the contract terms?"
Good answer: A clear answer in writing.
Warning sign: Pressure to sign for a long term before any proof.
8.Ask: "What will you not do or promise?"
Good answer: "No guarantees".
Warning sign: Any promise of a ranking or a recommendation.
Test it yourself in 30 minutes
- Write 5 questions your customers would ask an AI tool, in their words.
- Use a VPN set to your customers' country. Open a new chat in ChatGPT with search on.
- Ask each question three or four times. Write down who is named each time.
- Open ChatGPT's "Sources" and note which pages it relies on.
- Ask the same kind of question for the agency's client niche. Is that client named? If not, ask the agency why.
What we saw doing this ourselves (ChatGPT, search on, October 14, 2026, one run per question, so treat these as examples, not findings): asking the small-business agency question from a US IP instead of an overseas one changed the answer completely; for "Who can help my business get recommended by ChatGPT?" ChatGPT named no agencies and described the kind of professional to hire instead (when we first asked on September 30 it also named none); and for "best agencies" questions it mostly cited roundup articles.
Three promises that should end the meeting
- A guaranteed ranking, citation or recommendation.
- "An llms.txt file or schema will fix it." Google says Google Search doesn't use llms.txt and that structured data isn't required for its generative AI search.[1] More in What is llms.txt, and do you need it?
- Results from a single screenshot.
You may not need an agency yet
Do the free steps first: make sure AI's crawler can read your site, state plainly what you do, keep your facts consistent, and earn real mentions. Start with How do I get my small business recommended by ChatGPT? and Why doesn't ChatGPT mention my business? If you've done those and still aren't named, an agency may help.
The scorecard
| # | Criterion | 0 (missing) | 1 (partial) | 2 (clear, with proof) |
|---|---|---|---|---|
| 1 | Dated before-and-after proof | None | Charts or tool scores only | Dated screenshots, repeated runs |
| 2 | Questions tracked | Keywords only | Their list | Your buyer questions, you approve |
| 3 | Measurement | One screenshot | A few runs | Repeated runs, reported as how often you're named |
| 4 | What they'll change | "AI optimization" | General list | Specific pages, facts, listings |
| 5 | Where mentions come from | Won't say | Vague | Real reviews, directories, earned; no paid links |
| 6 | Who does the work | Vague "team" | Some names | Named people, one contact |
| 7 | If nothing moves | No answer | Verbal only | In writing, clear terms |
| 8 | What they won't promise | Promises results | Hedged | Clear "no guarantees" |
Swipe sideways to see every column.
Add up the points. We'd rather you use it on us than skip it.
Common questions.
Ask for dated proof from similar clients, check how it measures (repeated runs, not one screenshot), ask exactly what it will change, and test the agency's claims yourself using the 30-minute method above.
It depends on whether your customers ask AI tools for recommendations. Test first: if AI already names you, you may only need small fixes; if it doesn't, work out why (see the Why doesn't ChatGPT mention my business? guide).
No. Answers vary between runs, models and over time (SparkToro), and Google says indexing and serving aren't guaranteed (Google's guide).
Engines that search the web while answering can change within weeks once your pages and third-party mentions are in place. Answers from a model's built-in knowledge change only when the model is updated, which takes months. Nobody publishes reliable timelines.
No. A tool measures; an agency changes things. Some agencies resell a tool and call that the service, so ask question 4.
Sources
- Google Search Central, "Google's Guide to Optimizing for Generative AI Features on Google Search", last updated July 10, 2026.
- SparkToro, "NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibility", January 28, 2026.
- Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024 (arXiv preprint, 2023).
Not sure where you stand?
If you'd rather have it done for you, see our generative engine optimization services.