AEO and AI visibilityHow to Choose a GEO Agency: 12 Questions to Ask Before You Sign

A GEO agency works to raise how often AI engines like ChatGPT, Perplexity, Gemini and Copilot name and cite your brand when buyers ask questions. The same service is sold as an AEO agency, an AI visibility agency or a generative engine optimization agency. The labels overlap, which my comparison of AEO, GEO and SEO untangles.

Strip away the label and good GEO services look a lot like disciplined SEO and content work, pointed at a new unit of success. Google says plainly in its guidance on AI features and your website that there are no additional requirements to appear in AI Overviews or AI Mode.

Each engine finds its sources its own way, so the engine list should follow where your buyers research. ChatGPT search depends on OAI-SearchBot, Copilot citations show up in Bing Webmaster Tools, and Google builds its AI features on its core ranking systems. A GEO agency for B2B SaaS should test the engines your buyers use, not every engine that exists.

So the deliverables matter more than the name on the proposal. These are the ones I would expect to see in writing.

• A prompt set and test protocol. A small set of buyer questions, plus the engines, modes, location and account state used to test them.
• A dated baseline. Mention and citation rates by engine before any fix ships, with the competitors and sources cited instead.
• A crawler access check. Robots.txt and CDN rules for the AI search crawlers, and a check that answers exist in the raw HTML.
• A structural audit. Schema, answer-first headings, bylines, dates and internal links, scored page by page.
• The page-alone test. A model answers a flagship question from one page alone, which exposes pages that only imply their answer.
• Fixes and earned mentions. Page rewrites, new answer pages, reviews and digital PR that describes your brand the same way across third-party sites, ordered cheapest first.
• Reporting from first-party data. Your own Bing and Search Console AI reports, next to the agency's sampled tests.

The audit half of that list is the subject of my one-day AI visibility audit.

Your current SEO agency can handle GEO if it can show you a citation method, not just a new service page. Google's May 2026 guide to optimizing for generative AI features says that work is still SEO. In my book I argue the SEO function should own AI citation, next to search visibility reporting.

The risk is a keyword retainer with a new name. A specialist generative engine optimization agency earns the switch only when it brings a measurement discipline your current team lacks.

I would use five questions to tell the two apart. A team that already does this work answers them in a minute.

1. What is your test protocol? Location, language, account state and engine mode, written down and held fixed for a quarter.
2. How do you separate a mention from a citation? A citation is your URL in the sources list, and a brand name without a link is logged separately.
3. Do you report each engine on its own? Averaging a Bing win with a Perplexity loss into one score hides both.
4. How do you treat OAI-SearchBot and GPTBot? Search visibility and model training are separate decisions.
5. How do you read Bing's grounding queries? Filter on the query text, because the report's topic tags mislabel words with two meanings.

If the answers are vague, keep the agency for classic SEO and bring in the AI layer separately. My guide to AI SEO covers how the two fit together.

Ask about method, data ownership and guarantees, in that order. A strong GEO agency answers with specifics it can show you, while a weak one answers with screenshots and confidence.

Google's own advice on hiring an SEO is a good starting point, because the same buying mistakes repeat. It suggests asking whether a provider will share all the changes they make to your site, and it warns against anyone who guarantees rankings. I adapted that list for AI search and turned it into the scorecard below.

Score each answer as strong, partial or weak, and read the weak ones first. Two red flags on data ownership or guarantees should end the conversation, however good the rest sounds.

Ask for a case study with its method attached, and a short live test on your own brand. A before-and-after chart with no prompt set behind it proves little, because AI answers change from run to run.

A credible case study names the number of prompts, the engines and modes, the test protocol, the baseline date and the retest variance. It reports citations and mentions separately, by engine, and says what else changed in the same period.

Then ask for a sample audit before you sign. Give the agency three buyer questions and one priority page, and ask them to run the page-alone test while you watch. Where the model cannot answer or invents something, you see their diagnosis live.

Hold the reporting to the same standard. Each metric below answers one question, and none of them proves revenue on its own.

Be honest about scale. At the data and AI services company where I run this workflow, 677 visitors arrived straight from ChatGPT, Perplexity, Gemini and Copilot answers in 18 months. That is small, and it is still why my book calls this the most time-sensitive workflow, because the patterns engines learn now carry inertia.

Apply one more filter to any list of generative engine optimization companies or top answer engine optimization companies. Check who published it, because an agency ranking agencies has a stake in the order. My guide on how to measure AI visibility goes deeper on reading these numbers.

Walk away from guarantees, secret markup and data you cannot take with you. The rest of this list is negotiable, but each one tells you how the agency thinks.

• Guaranteed citations or rankings. Google's hiring advice says no one can guarantee a #1 ranking, and AI answers are less stable still. SparkToro's January 2026 research ran 12 prompts a combined 2,961 times and found ChatGPT and Google's AI had less than a 1 in 100 chance of repeating the same brand list.
• Secret markup and Google hacks. Google says you need no new machine-readable files, AI text files or special schema for its AI features. Its May 2026 guide adds that you can ignore chunking content, unnecessary llms.txt files and inauthentic mentions for Google Search, and my llms.txt guide covers where the file still has a use.
• One engine only. An agency that only talks about ChatGPT, or only about Google, will miss part of where your buyers research.
• Keyword stuffing for AI. In the original GEO paper, keyword stuffing scored below the unoptimized baseline. The authors called it ineffective for generative engines.
• Mass-produced AI pages. Google's spam policies name using generative AI to create many pages without adding value as scaled content abuse.
• Blocking the wrong crawler. OpenAI's crawler documentation says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. Blocking GPTBot only opts you out of model training.
• Owning your data. If the prompt set, logs and dashboards sit only in their account, you are renting your own baseline.
• Long lock-in before a baseline. A twelve-month minimum signed before anything is measured means paying for a year to learn whether the method works.

None of these red flags is about agency size or price. A small shop can fail every one, and a large one can pass them all.

GEO agency pricing depends on scope, so compare what each month buys rather than the fee. One marketing talent marketplace's June 2026 pricing guide runs from $3,000 a month for a lean retainer to $20,000 or more for full-service work at enterprise scale, with one-time audits starting around $5,000.

Most proposals use one of four structures.

• Fixed-scope audit. A baseline, a crawler and structure review and a ranked fix list. It is the cheapest way to see an agency's method.
• Monthly retainer. Ongoing tracking, fixes, content and earned mentions. Ask for the deliverables per month in writing.
• Paid pilot. Thirty days on a fixed scope, with a decision at the end. The pilot section below shows how I would run one.
• Retainer plus performance fee. Part of the fee tied to agreed targets. Tie it to logged citation rates on a fixed prompt set, never to screenshots.

The price moves with the number of engines and markets, the size of the prompt set, content volume, digital PR scope and the technical work your site needs. A small flagship set tested well costs less and tells you more than a large set tested once.

The tracking half is the cheapest part to keep in-house, because it is a fairly easy workflow to build. Pay an agency for judgment, fixes and earned coverage, and track AI referral visits and the pipeline they touch in your own CRM.

Write ownership, data access and exit into the contract before the work starts. Pitches change after signature, and a contract does not.

1. You own the prompt set and the logs. The flagship questions, the test protocol and every citation log row belong to you, during the engagement and after it.
2. A monthly raw data export. Every run, engine, prompt, date, mention, citation and source cited, in a spreadsheet you can open without their tool.
3. A named owner. One senior person accountable for the program, named in the contract, with notice if that changes.
4. A baseline before fixes. A dated baseline under the written protocol, delivered before any change ships, so later gains are measured against something real.
5. A change log. Every edit to your site, robots.txt, CDN rules or schema, with the date and the reason.
6. Your accounts, their access. Bing Webmaster Tools, Search Console and analytics stay in your name, with user access granted to the agency.
7. A clean cancellation. A short notice period and a handover pack with the logs, the protocol and the open fix list.

Two first-party sources make the data clause easy to enforce. Bing's AI Performance report shows citations across Copilot and Bing's AI summaries, and Google's generative AI performance reports in Search Console reached all websites on August 31, 2026. Neither covers ChatGPT or Perplexity, which is why the sampled test log still matters.

Pick the option that leaves the most capability behind for the budget you have.

Whoever does the work, someone inside has to answer for the citation log. My advice on agencies in the book is short. Keep the agency if their expertise earns it, and make sure what they find becomes yours.

I work as a fractional consultant, so weigh my view of the third row with that in mind. A consultant builds and teaches while an agency can supply more hands, and a good agency on a strong contract can leave you just as much behind. If you go tool only, my comparison of AI visibility tools covers what to check before you subscribe.

Use the first 30 days to test the method, not to judge citation gains. Engines re-index slowly, so a month rarely proves lift, but it shows clearly whether the agency measures honestly and fixes the right things first.

1. Week 1, baseline. Agree on 10 to 15 flagship buyer questions, two control prompts you cannot win and one canary you already win. Write the test protocol, run the baseline, and export your Bing and Search Console AI data.
2. Week 2, audit. Check crawler access and raw HTML, score the priority pages, and run the page-alone test. The output is a fix list ranked cheapest first, with restructuring a thin page ahead of writing a new one.
3. Week 3, cheap fixes. Ship three to five low-cost fixes, each one in the change log. OpenAI notes a robots.txt change can take about 24 hours to reach ChatGPT search, so do not retest the same day.
4. Week 4, rerun and review. Rerun the full set under the same protocol and retest 10% of the question and engine pairs in a second clean session. If more than 20% disagree, the method needs fixing before anyone reads the results.

Pass the pilot if the baseline is reproducible, the data arrived without chasing, and the fix list made sense to your own team. Fail it if the numbers only exist in their dashboard, the protocol changed mid-month, or the canary failed and nobody flagged it.

A pilot also tells you how the program should fit your wider plan. My AEO strategy guide covers where it sits next to SEO and content.

This week, write the three questions a serious buyer asks before contacting you. Run each one through ChatGPT, Perplexity and Copilot, and save all nine answers with the date and the sources cited.

That small file becomes your baseline. Any GEO agency you talk to should beat it, explain it and add to it. If you would rather build the program inside your team so you keep it, see how I set up AEO and GEO programs your team owns.