What to tell a client who asks about AI search

The question usually arrives with anxiety attached. A client saw something on a podcast, or a competitor's agency sent a scary email, and now they want to know why AI search is not on your invoice.

The instinct is to add a line item so you look current. Resist it. The stronger position is to be the person in their life who explains the thing accurately, because that is worth more than a retainer increase and it lasts longer.

What do you say in the first two minutes?

Three sentences, then stop talking.

AI answers are assembled from search results, so being findable in search is being findable in AI. Google's own documentation puts it this way: "From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO." The parts that are new are how we measure it and how directly we answer questions on the page.

If they push, add the technical requirement, which is two conditions and no more. The page must be indexed, and it must be eligible to be shown with a snippet. Google states there are no additional technical requirements. Not a file, not a markup type, not a format.

Then let them ask. They usually ask one of three things: are we in the answers, do we need to do something different, and is our competitor doing something we are not.

The honest replies are: sometimes, and it varies run to run. Mostly no, with two exceptions. And probably not, but let us look.

The client is not asking for a strategy. They are asking whether they are being left behind. Answer that question, not the technical one.

What if they send you a study with a big number in it?

Trace it in front of them. The most common one traces in about five minutes and the trace is more persuasive than any opinion you can offer.

The claim, as it arrives. "GEO delivers 40% more AI visibility, and it is peer-reviewed."

Where the number comes from. There is a real paper. "GEO: Generative Engine Optimization," accepted to KDD 2024. Table 1 reports a No Optimization baseline of 19.3 and a best-performing tactic, Quotation Addition, at 27.2. Divide one by the other: 27.2 / 19.3 = 41%derived. That is the whole provenance of the figure in the email.

What the 19.3 and the 27.2 measure. A metric called position-adjusted word count, meaning how much of a generated answer's text is attributable to your source, weighted by where it appears. That is not traffic, not citations, not customers, and not a number your client can bank.

How it was produced. Not from a live product. The researchers fetched the top five Google results for a query, fed the cleaned text to GPT-3.5-turbo with a fixed prompt, and sampled the output five times. It is a simulation of a generative engine over a fixed candidate set.

What was in the benchmark. 10,000 queries, 80% of them informational, across 25 domains. Nothing in it is a local commercial query. The paper has never been validated for local intent.

And the effect is not portable. The same tactic ranged from 19.1% to 35.8% in effectiveness depending on the subject category. There is no universal number to promise.

So price the proposal against that. If the sender wants $1,800 a month, $21,600 a year, for GEO work justified by that figure, the honest sentence is: the cited lift is on a word-count metric, in a simulation, on queries unlike yours, with a category range wide enough that nobody can tell you your number.

Here is the part that keeps you honest in the other direction. The paper's findings are real and some of them are useful. Adding direct quotations, verifiable statistics and inline citations all scored above the baseline. Keyword stuffing scored 17.7, below the baseline, meaning it actively hurt, and an authoritative tone produced no significant improvement. Those are free writing habits, not a retainer. Give the client the habits and skip the invoice.

What if they have a competitor's proposal in hand?

Read it with them, line by line, and sort it into three piles.

Pile one, work they already pay you for under a different name. Technical crawlability, content, digital PR, profile management, reviews. Real work, renamed.

Pile two, genuinely new work. Prompt-set sampling with stated sample sizes, crawler log analysis verified by IP range, and rewriting key pages to answer in the first three sentences. Small pile, honest pile.

Pile three, things Google names as ineffective: llms.txt and similar special files, artificial content chunking, content rewritten for AI systems, and seeking inauthentic mentions.

Pile three is your whole argument, and it needs evidence rather than opinion. On the file specifically, a study of roughly 137,000 sites found 97% of published llms.txt files were never fetched by anything at all. That number ends the conversation more politely than any amount of professional disagreement.

Google supplies one more sentence for the tooling layer, and it is worth reading aloud: "Be wary of third-party tools that promise ranking success or claim to use 'internal' Google metrics. No third-party tool has access to our internal ranking or AI systems."

Do not sneer at the competitor. Sort the document. The sorting speaks for itself and you come out of it as the adult.

What if they ask for a number?

Tell them a single AI visibility score is not a real number, then give them something better.

These systems are non-deterministic. A University of St. Gallen team ran four assistants daily for a 45 day window and also ran identical prompts ten times inside a single day. Between two consecutive days, only 34% to 42% of cited sources overlapped. Same day, same prompt, 32% to 43%. There is no position to report because the source set behind the answer turns over that fast.

What you can report is a rate: named in seven of twenty runs of this exact prompt, on this date, from this location, with the run count printed next to it. The same paper prescribes at least seven runs per prompt per day and a two to four week rolling window rather than week over week reporting, which is a method you can put in a scope of work and defend.

Clients accept that more readily than agencies expect, because a number with a method attached feels more solid than a score with none. Non-determinism is the mechanism underneath all of it and it is worth being able to explain in one breath.

Then show them who else is getting named. Which sources AI assistants actually cite for local queries turns an anxious question into a work list, and a work list is what they wanted from the meeting anyway.

What should they be worried about instead?

The documented local trends, which are less exciting and considerably more expensive.

Sterling Sky's 2026 field data shows clicks-to-call from Business Profiles declining even for profiles holding stable rankings, and local pack ads going from roughly 1% of tracked mobile reports in early 2025 to roughly 22% by December 2025. Joy Hawkins's summary is that Google is going pay-to-play. That is vendor field data from a local SEO agency, with the tooling disclosed.

Rank the urgency honestly for them. A disappearing call button and a results page filling with ads are happening now, to them, in their category. AI search is a hedge against a shift that is real but early. A client who reprioritizes a quarter around AI visibility while their call volume quietly erodes has been badly advised, even if every AI sentence in the advice was true.

What do you commit to doing?

Five things, and they should already be familiar.

Make sure every important page is indexed and eligible for a snippet. That is the entire technical requirement Google states.

Rewrite the top pages to answer their question in the first three sentences and keep paragraphs self-contained.

Get the Business Profile exactly right. Categories carry the most weight, and picking the primary category when several fit is a decision worth spending real time on.

Keep the reviews coming, permanently. And say the compliance line precisely, because most people get it wrong. Routing happy customers to a public review link while diverting unhappy ones to a private form is explicitly banned by Google, in the clause prohibiting merchants from discouraging negative reviews or selectively soliciting positive ones. The FTC's 2024 rule does not name gating by that word; it reaches the adjacent conduct, meaning incentives conditioned on sentiment and misrepresenting a filtered set of testimonials as representative. Saying "the FTC bans gating" is the kind of imprecision a client's lawyer will correct. Gating is prohibited and it also makes you rank worse is the version clients remember.

Measure with prompt sampling and server logs, and report both with methods attached.

One more decision to raise before they hear it elsewhere: whether to allow AI crawlers at all. For a local business the answer is almost always yes, and the case for blocking is narrower than it sounds.

What to do this week

1. Write your three-sentence answer and use the same one every time. Consistency reads as competence. 2. Read Google's AI optimization guidance yourself. Fifteen minutes, and it changes every conversation you have afterwards. 3. Build a prompt set for each client in their customers' language. Twenty sentences, not keywords. 4. Run five prompts seven times for your largest client and put the rates in the next report, with the method stated. 5. Add a crawler log line to reporting for any client whose logs you can access. 6. Audit your own proposal for anything in pile three. If it is there, take it out before someone else finds it.

Need the words for a specific client conversation?

Send me the client's question exactly as they asked it, at eric@seod.com, and I will send back a reply you can paste into an email, in plain language, with the documentation reference where one exists. If they forwarded you a study, send that too and I will trace the headline number back to its table. No charge, no pitch, and I will not use your client's name for anything.

I offer this because agency people get asked this question weekly and the honest answer is harder to write than the impressive one, and the honest answer is the one that keeps clients.

Otherwise, there is more on AI search and AEO here.