What Your SEO Agency Isn't Telling You About AI Search

Assuming your SEO agency has an AI search plan is a mistake. The retainer line item changed names last year, "AI visibility," "GEO add-on," "AEO coverage," the invoice went up 20 to 30 percent, and the monthly deck gained a new slide with ChatGPT in the title. In reality, a lot of what agencies are selling as generative engine optimization is improvisation under a new label, priced to look expert and reported to look measurable.

The distance between what agencies say they do for AI search and what they can actually prove is where the money leaks. This piece is a buyer-side audit: the specific omissions, the euphemisms on the deck, and the questions that force an honest answer out of your current provider.

The Repricing Nobody Explains on the Call

AI search did not create new work for most agencies. It created a new line item. Independent surveys now put the median GEO and AEO retainer at $6,850 per month, roughly 45 percent higher than the median traditional SEO retainer, with 68 percent of agencies pricing it as a monthly retainer rather than a defined scope of work. Playbooks aimed at agency owners are explicit about the move: add GEO to an existing SEO retainer at a 20 to 30 percent uplift on current MRR.

None of that is inherently dishonest. The dishonesty is in the omission. The uplift rarely comes with a dated change-of-scope document, a list of removed deliverables, or a new definition of success. You get the higher invoice and the same monthly rank report with two AI slides stapled to the back.

Ask for the delta in writing. If your retainer rose and the new line is "AI visibility," the proposal should name the surfaces being optimized, the queries being tracked, and the specific artifacts (schema, entity pages, citation-earning assets) being produced each month. If none of that exists, you are paying rent on a vocabulary change. The task-level differences between SEO and GEO work are real; the invoice should reflect them.

ChatGPT Citations Are Rarer Than the Dashboard Suggests

The second quiet omission is how often AI answers cite anyone at all. Similarweb's analysis found that 6.8% of U.S. ChatGPT responses carried citations in May 2026, up from 1.6% a year earlier. Four-fold growth is real, but the base is tiny. The overwhelming majority of ChatGPT answers resolve without surfacing a source URL at all.

That matters for how your agency reports ChatGPT citations. A "citation share" number pulled from a 9-million-query tracker is not the same as a share of answers your buyers actually see. If the agency is reporting a 3 percent citation rate inside a universe of queries that themselves only cite 6.8 percent of the time, they are reporting a share of a share. Ask what the denominator is. Ask whether the queries being tracked are ones real customers run, or ones chosen because your brand already shows up for them.

The other half of this problem is retrieval. One study of 548,534 pages across 15,000 prompts, found that ChatGPT cited only 15% of what it retrieved. Being read by the model and being cited by the model are different outcomes, and most GEO reports conflate them.

Single-Platform Reporting Hides Most of the Market

Agencies tend to over-index on whichever platform is easiest to track. In practice, that is ChatGPT. Statcounter data shows ChatGPT sent 79.4% of AI chatbot referrals worldwide in August 2026, with Gemini a distant second at 10.9 percent, Perplexity at 4.3 percent, Copilot at 2.8 percent, and Claude at 2.6 percent. ChatGPT dominance in referral traffic is real, and it does justify prioritizing it.

But referral traffic is only one measure of exposure. Session share tells a different story: Gemini holds 27.7% of AI chatbot sessions and Claude has climbed to 10.3% on a session basis, even though both send far less outbound traffic. A brand with strong ChatGPT citations and nothing on Gemini is invisible to roughly a quarter of the AI search market by session volume. There is also almost no overlap between what each engine cites. Averi's analysis across hundreds of millions of citations finds ChatGPT and Perplexity share about 11 percent of the domains they cite. Single-engine measurement hides most of the picture.

If your monthly report shows ChatGPT visibility and nothing else, that is a reporting constraint, not a strategy. Ask which engines are tracked, how often, and against what prompt set. If the answer involves one tool and one platform, you have an incomplete read. The specific metrics worth watching on Claude look nothing like the ones that matter on ChatGPT.

The llms.txt Theater

A lot of agencies added "llms.txt setup" to the deliverable list in 2025. The honest status of that file is now awkward for them. Google's May 2026 generative AI search guidance explicitly tells site owners they don't need to create new machine-readable files, AI text files, markup, or Markdown such as llms.txt to appear in Google Search or its generative AI features. An Ahrefs server-log study of roughly 137,000 domains found that 97% of llms.txt files received zero requests from AI crawlers during the study month.

That is not a reason to never ship one. It is a reason to stop invoicing it as a visibility lever. If llms.txt is on your deliverable list, it belongs in the housekeeping tier, not the strategy tier. A fair question for your agency: which AI crawlers have actually hit our llms.txt file in the last 30 days, and from which log source are they measuring it? If no one can answer, the file is decoration. For more on how practitioners are reading the data, see our take on llms.txt adoption in 2026.

The Reporting Theater Around AI Traffic

The last omission is the most consequential, because it swings in the client's favor and most agencies still don't surface it. AI referral traffic is small in volume but disproportionate in value, and it has flipped on conversion in a year.

Adobe Analytics data showed AI-referred traffic to U.S. retail sites converted 42% better than non-AI traffic in March 2026, a reversal from March 2025 when AI traffic converted 38 percent worse than non-AI channels. By July 2026, AI-referral traffic generated 53% more revenue per visit than non-AI traffic across the same panel. The randomized field experiment on AI Overviews showed Google's AI summaries reduced outbound organic clicks by 39.8%, so volume is down, but the volume that remains is qualified differently.

Any agency reporting that still leads with sessions, impressions, and keyword rank is reporting the wrong thing. The relevant question is revenue per AI-sourced visit, segmented by engine. If your agency cannot show that number, they are not measuring the channel they are charging you for. Our longer argument on why lower organic traffic can be the right outcome walks through what to measure instead.

Questions That Force an Honest Answer

Run these on your current retainer. The quality of the answers tells you whether you have an AI search strategy or a repricing.

• What changed in the scope when the retainer went up? Ask for the before-and-after deliverable list, dated. "AI visibility" is not a deliverable.
• What is the prompt set you track, and who chose it? If the agency picked the prompts, they picked ones where you already win. Buyer-chosen prompts are the honest set.
• Which engines do you measure, and how? ChatGPT via a scraping tool is not multi-engine coverage. Ask about Gemini, Perplexity, Claude, and AI Overviews separately.
• What is the citation rate inside the universe you report? If they cannot name a denominator, the share number is cosmetic.
• Show me revenue per AI-sourced visit for the last 90 days. This is the number Adobe has published; your agency should be able to produce the equivalent from your analytics.
• Which crawlers have hit our site and our llms.txt in the last 30 days? A log answer, not a slide answer.

Several of these overlap with the broader diligence questions worth asking any SEO company before you sign, and all of them apply equally when the service is being rebadged as GEO or AEO.

Where This Leaves the Retainer

The point is not that AI search is unknowable or that every agency is bluffing. Google's own guidance, independent server logs, and conversion panels now give buyers enough ground truth to tell measurement from theater. The point is that the honesty asymmetry has shifted. Three years ago an SEO agency could show rank reports and the client had no competing read. Today the client can ask for log data, revenue per visit, engine-level citation rates, and dated scope changes, and the answers are either on hand or they are not.

Agencies improvising on AI search tend to answer in vocabulary. Agencies doing the work tend to answer in artifacts. The questions above sort them.