Service

AI Search Optimization
AEO and GEO

Updated 3 August 2026 · By Mark Gabrielli, Fractional CMO

Why this is a budget line in 2026 and was not in 2024

The honest version of the market case does not need inflated numbers. Gartner's 2026 CMO Spend Survey, published 11 May 2026 and fielded across 401 marketing leaders between January and March, puts AI at 15.3% of marketing budget. That is the part most people quote. The more useful pair of figures sits next to it: 70% call AI a key focus for the year, and 30% say they have the infrastructure to act on it.

Source: Gartner 2026 CMO Spend Survey, published 11 May 2026, n=401, fielded January to March 2026. Gartner's sample skews to large enterprises, so read these as directional for a company under $50M in revenue. The same figures and that caveat are carried on our marketing statistics page.

That 40-point gap between intent and capability is the actual argument for doing this now. It is unusual for a channel to be simultaneously acknowledged as important by most of a market and executed by a minority of it. That window closes.

What actually gets a company cited

There is a lot of noise in this category, most of it repackaged SEO advice. In practice the work that moves citation probability falls into four buckets, and only one of them is content in the way marketers usually mean it.

Entity clarity. A model that cannot confidently resolve who you are will not put your name in an answer. That means one consistent description of the company, consistent naming across every profile that mentions it, an unambiguous relationship between the company and the people who speak for it, and structured data that says the same thing the prose says. Most of the failures we see are here, and they are unglamorous to fix.

Answer-first structure. Answer engines lift passages, not pages. A page that buries its answer under 400 words of preamble is a page whose answer does not get lifted. Every question a buyer asks should be answerable in the first 40 to 120 words under the heading that asks it, with the supporting argument after, not before.

Structured, comparable proof. Models reach for content that is already in the shape of an answer: a dated table, a labelled range, a specification with units. Vague qualitative copy loses to a table with a date on it almost every time, which is why our own money pages carry dated figures rather than adjectives.

Third-party corroboration. The single hardest lever, and the one no vendor can sell you as a subscription. A claim that appears only on your own domain is a claim a model has no reason to trust. The same claim appearing in an industry publication, a directory with editorial standards, or a customer's own write-up is a different object entirely.

The AI crawlers that decide whether you get cited, and what each one actually does

Most published crawler lists are copied from other lists, and they drift. The table below was rebuilt on 19 August 2026 by reading each vendor's own documentation rather than a summary of it. Two rows are the reason it is worth having. First, the user-triggered fetchers are a real gap in most robots.txt files: OpenAI and Perplexity both state plainly that when a person asks a question and the product goes and reads a page, robots.txt is generally not applied. If your reason for blocking a bot is that you do not want your content in an answer, robots.txt does not close that door, and only a server-level or WAF rule does. Anthropic is the exception on that row and says its user fetcher does respect robots.txt. Second, several tokens in wide circulation are not real. There is no Gemini-Bot in Google's documentation; the token that governs Gemini training is Google-Extended.

Sources, all read directly on 19 August 2026: OpenAI bot documentation (developers.openai.com), Anthropic crawler support article (support.claude.com), Perplexity bot guide (docs.perplexity.ai), and Google's common crawlers reference (developers.google.com). The token, operator, purpose and robots.txt columns are the vendors' own statements. The recommendation column is our judgement and is not part of any vendor documentation. Crawler tokens change; treat any list older than a quarter as unverified, including this one after 19 November 2026.

How this differs from SEO, precisely

The technical foundations overlap almost completely. If a page cannot be crawled, rendered and parsed, none of this matters, which is why AI search work that skips the technical audit is theatre. The difference is the unit of optimization. SEO optimizes a page against a query and competes for a ranked position. AI search optimizes a passage against a question, and behind the passage it optimizes the entity, because the model has to decide not only that the answer is good but that you are a credible thing to name.

The second difference is commercial and it is the one that catches finance teams off guard. There is frequently no click. A buyer can read your position, your pricing logic and your differentiation inside an assistant's answer and arrive at your site already decided, or never arrive at all and still buy. Attribution models built on last-click will report this channel as producing nothing while it is quietly moving your pipeline. If you are going to fund this, fix the measurement question before the first invoice, not after.

How we measure it, and what the measurement cannot do

We run prompt panels rather than rank trackers. A fixed set of real buyer questions is run against each assistant on a schedule, and every response is scored for three things: whether the brand is mentioned at all, whether it is cited with a link, and which specific page was pulled. Over time that gives a citation rate per assistant per question, and the page-level detail tells you which asset is doing the work so you can build more of that and less of everything else.

What it cannot do is give you the precision a rank tracker gives you. Responses are non-deterministic, so the same prompt run twice can differ. Retrieval sets change between model versions without notice. Sample sizes are small because each observation costs a call and a human read. Anyone presenting AI search results with the confidence of a rankings report is either not sampling properly or not telling you about the variance. We report the variance.

What an engagement includes

AI search optimization is rarely a standalone engagement here, and that is a deliberate position rather than a packaging preference. It sits inside a fractional CMO retainer because the work it depends on, positioning, proof, pricing clarity and the content operation, is CMO work. A programme typically covers an entity and schema audit across the properties that describe you, restructuring of the money pages into answer-first form, a dated-proof content programme aimed at the questions your buyers actually ask, an off-domain corroboration plan, and the prompt panel that tells you whether any of it worked.

Engagements run $8,000 to $25,000 per month depending on scope. For what that buys and how it compares to a full-time hire or an agency, see the fractional CMO cost breakdown.

Who runs this work

Mark Gabrielli is a fractional Chief Marketing Officer with 15+ years of B2B marketing leadership, $135M+ in marketing-attributed revenue from seed through Series B, 90% client retention, and a practice serving 370+ US cities across all 50 states. The AI search work is run the same way as the rest of the practice: a deliberately small book of engagements, figures published with their source and date, and claims that survive being checked. More on the background and the record is on the about page.

Frequently asked questions

What is AI search optimization?
It is the practice of making a company the source an AI assistant reaches for when someone asks a question in its category, covering both AEO for direct-answer surfaces and GEO for citation inside generated responses. The levers are entity clarity, answer-first structure, correct schema, and off-domain corroboration.

How is AI search optimization different from SEO?
SEO competes for a ranked position on a page of links. AI search competes to be the source inside a generated answer, where there is no position one and often no click. The technical foundations are shared; the unit of optimization is a passage and an entity rather than a page and a query.

How much does AI search optimization cost?
Inside a fractional CMO retainer, $8,000 to $25,000 per month depending on scope. As a standalone project the cost sits in content and engineering time, not tooling. Treat licence-fee pricing for AI search with suspicion: the durable work is labour.

Can you guarantee ChatGPT will cite my brand?
No, and a guarantee is a reason to walk. Model responses are non-deterministic and no vendor controls what a model emits. What can be influenced is whether your facts are unambiguous, structured and corroborated, which is what raises citation probability.

How do you measure AI search performance?
Prompt panels run on a schedule against each assistant, scored for mention, citation with link, and which page was pulled. It is coarser than rank tracking and the variance is reported rather than hidden.

How to verify an AI crawler actually visited you

A user agent string is a claim, not evidence. Anyone can send one. The only way to know that ChatGPT, Claude, Perplexity, Google or Copilot genuinely fetched a page is to match the request in your server logs against the IP ranges the operator publishes. Every operator on this list now publishes those ranges as a machine readable file. The table below was built by fetching each of those files on 2026-09-04.

Two things in that table matter more than the individual numbers, and neither is widely discussed.

First, the AI operators publish IPv4 only. Across the OpenAI, Anthropic and Perplexity files there were 0 IPv6 prefixes published on 2026-09-04. Google published 283 across its files. If your origin answers on IPv6 and you build an allowlist from these files, you have built an IPv4 allowlist and you should not fail closed on it. Verify, log, and look at what enforcement would have blocked before you enforce anything.

Second, the files age at completely different rates. Google regenerated all of its range files the day before this page was updated. Other files on the same list carry regeneration dates many months older. A list that has not moved is not necessarily wrong, but a verification job that assumes freshness is. Re-fetch on a schedule, keep the previous copy, and alert on a change in size rather than trusting whatever you cached.

What this method cannot tell you

This is a verification method, not a measurement of visibility, and it is worth saying where it stops. Confirming that a crawler fetched a page tells you the door was open and the page was taken. It does not tell you the page was used, that any answer quoted it, or that a single person saw your name. Those are the questions the measurement section above deals with, and they are answered by prompt panels and sampling, not by logs.

It also cuts against a common piece of advice. Access is a floor, not a lever. Once a crawler can reach the page, more crawler access does not produce more citation, and a great deal of what is currently sold as AI search optimization is work on that floor. Google's own documentation on its AI features states that you do not need to create new machine readable files, AI text files, or markup to appear in them, and that there is no special schema.org structured data you need to add. That is an uncomfortable sentence for this category and it is worth repeating honestly rather than working around.

Crawler verification questions, answered

How do I know an AI crawler actually visited my site?

Check your server access logs for the user agent, then verify the source IP against the operator's published range file. The user agent on its own is not evidence, because it is a string the client chooses and anyone can send it. Every major operator now publishes a machine readable list of the addresses its crawlers use, which is what turns a claimed visit into a confirmed one.

Can I just block anything that is not on the published IP lists?

Be careful. As of 2026-09-04 the AI operators publish IPv4 ranges only, while Google publishes both IPv4 and IPv6. If your origin accepts IPv6 and you fail closed on an IP allowlist, you can block traffic you meant to allow. Verify and log first, and only enforce once you can see what enforcement would have done.

How often do those IP lists change?

Unevenly, and that is the practical problem. Measured on 2026-09-04, Google's files had been regenerated the previous day while other files on this list had not been regenerated for many months. A verification job has to re-fetch on a schedule and treat a stale file as a reason to log rather than a reason to block.

Does an llms.txt file help me get cited?

There is no evidence that it does, and Google's own documentation on its AI features says plainly that you do not need to create new machine readable files, AI text files, or markup to appear in them, and that there is no special schema.org structured data you need to add. Publishing one is cheap and harmless. Treating it as a ranking or citation lever is not supported.

If a crawler fetched my page, does that mean I will be cited?

No. Crawl access is necessary and nowhere near sufficient. Verifying a fetch tells you the door was open and the page was taken. Whether the page was then selected, quoted or linked in an answer is a separate question that the logs cannot answer, and it is the question the measurement section above deals with.

AI search optimization by state

State and metro pages covering local market context: Texas California Florida New York Illinois Georgia Colorado Washington Massachusetts North Carolina Arizona Pennsylvania and every other US state.