AI SEO consulting — what you are buying when there is no checklist

Google publishes no AI-search checklist, so what is there to buy?

A page selling AI search consulting can imply a proprietary method. This one starts by ruling one out, because the platform has already said it does not exist. Google’s guidance for site owners on AI features states there are “no additional requirements to appear in AI Overviews or AI Mode”, and no other special optimisations are necessary. The same document says you do not need to create new machine-readable files, AI text files or markup to be eligible, and that there is no special schema.org structured data to add. It also says that meeting every requirement does not oblige Google to crawl, index or serve a page — indexing and serving are not guaranteed. Eligibility to appear as a supporting link is the ordinary bar: the page has to be indexed and eligible to show with a snippet.

If that is true — and it is the platform’s own statement about its own product — then no consultant can honestly sell you an AI-search tactic list, because there is no list to sell. What remains scarce, and what this engagement actually is, comes down to three things: deciding which of several genuinely different problems is the one blocking you, deciding the order in which to spend against it, and deciding in advance how the result will be counted. None of those are tactics. All three are judgement calls that go wrong expensively.

One related piece of folklore is worth removing at the same time. The Search Quality Rater Guidelines overview, Google’s own explainer for the programme, describes a rating instrument that human evaluators apply to search results. The guidelines are not a list of ranking factors and they are not an AI-search specification. We will point at the actual text when a question is genuinely about how experience and expertise are assessed. We will not present that document as a checklist, because it does not work as one, and treating it as one is how agencies end up billing for “E-E-A-T optimisation” with nothing checkable underneath it.

Where this engagement stops and our AEO, GEO and AI Search SEO pages start

This is a consulting engagement, not a fourth synonym for the same work. The boundary is worth stating exactly, because the three terms are easy to confuse, and confusing them can mean paying twice.

• AEO is a property of one page. It asks whether a machine can lift a clean answer out of a specific page of yours. That is structural, and it is explained in full on our AEO services page.
• GEO is a property of the web’s agreement about you. It asks whether an engine composing an answer has a reason to name you, and part of the answer to that sits outside your own website. The mechanism is set out on our GEO services page.
• AI Search SEO is the question of which of those is your binding constraint. The comparison, the symptoms and the decision tree live on our AI Search SEO page, free to read, with nothing gated.

Those three pages are diagnostic and educational. This engagement is what happens after the diagnosis: someone runs it against your actual site and your actual query set, argues a view about which single constraint is binding, sequences the work against that view, and commits in advance to what will and will not be measured.

The honest corollary is that a lot of readers do not need to hire anyone. If you read the AI Search SEO page, can already tell which of the three describes you, and have people who can execute, buy the execution and skip the consulting. If the blocker is underneath all of it — crawling, rendering, indexation — that is technical SEO, and no amount of strategy substitutes for it. We would rather say that at the enquiry stage than sell a diagnosis you already have.

Decision one: your AI crawler policy, decided rather than defaulted

This decision is commercial and legal before it is technical, and in our census a robots.txt block on AI agents was uncommon: only 121 of 1,948 reachable Wikidata-listed Indian business hosts had a root-level robots.txt rule blocking any of the four AI agents tested.

OpenAI documents four separate agents with four separate jobs: OAI-SearchBot, which indexes sites for ChatGPT’s search features; GPTBot, which crawls content that may be used to train foundation models; OAI-AdsBot, which checks pages submitted as ads; and ChatGPT-User, a fetch triggered by an individual user’s request rather than by automated crawling. OpenAI states that each robots.txt setting is independent of the others, and gives the example of allowing the search agent while disallowing the training agent. Permitting search while declining training is therefore a documented, supported position, not a workaround.

Four things follow that are easy to get wrong:

1. Blocking ChatGPT-User is not a search opt-out. OpenAI documents it as user-initiated, notes that robots.txt rules may not apply to it, and says explicitly that it is not what determines whether content can appear in search. OAI-SearchBot is the control for that.
2. Disallowing the search agent is narrower than disappearing. OpenAI says sites opted out of OAI-SearchBot are not shown in ChatGPT search answers but can still surface as navigational links. The cost of the decision is specific, and worth stating specifically rather than catastrophising it.
3. The change is not instant. OpenAI notes that, for search results, it can take roughly 24 hours from a robots.txt update for its systems to adjust. A policy change and a measurement window must therefore not overlap, or you will attribute a reporting artefact to your own work.
4. The block may not be in robots.txt at all. Bot management at the CDN or WAF layer can stop agents before the file is ever consulted, and those rules do not appear in robots.txt. Editing robots.txt in that situation changes nothing and produces a plan built on a false premise.

Our own census gives the base rate. The AI Search Readiness Benchmark reached 1,948 Wikidata-listed Indian business hosts on 3 September 2026. Four of them — 0.2% — disallow OpenAI’s search crawler. Ninety-eight block OpenAI’s training crawler while permitting its search crawler, and none do the reverse. So blocking OpenAI’s training crawler while leaving its search crawler open is roughly twenty-five times more common than blocking both. The census records the rule, not the reason for it: it shows that 98 of the 121 hosts blocking any of the four agents tested have this split, not that each of those 98 hosts weighed the choice. On those numbers a robots.txt block on search is rarely the reason a business is missing from AI answers — though the census read root-level robots.txt only, so it cannot see the CDN and WAF rules described above.

What the engagement adds is that the position gets chosen rather than inherited, by someone with the authority to choose it. Legal and communications may have a view on training. Marketing has a view on search. Those two views can differ and still be coherent, which is the entire point of the settings being independent — but only if somebody writes the decision down, dates it, and confirms it was applied at the layer that actually enforces it.

Decision two: what will be counted, and what genuinely cannot be

Google reports AI-feature appearances inside overall search traffic. Its documentation states that sites appearing in AI Overviews and AI Mode are included in the Search Console Performance report within the Web search type — not broken out as a separate AI channel. Google’s documentation describes no filter that isolates AI-feature traffic from the rest.

That single constraint decides what an honest measurement commitment can look like, and it has four consequences.

• An “AI traffic” line item is modelled, not measured. If a provider shows you one, the useful question is which report and which filter produced it. If the answer is a proprietary dashboard, you are being shown an estimate with an unstated error bar, presented with the visual authority of a measurement.
• Assistant answers are samples, not metrics. The same prompt can return different sources by session, region and model version, and a small change of phrasing changes them further. A stable-looking “AI visibility score” implies a stability that the underlying system does not have. What can be reported honestly is a sample carrying its prompt, its date, its region and the model version, plus the direction of travel across several such samples.
• Query fan-out breaks keyword-level thinking. Google documents that AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources to compose a single response. A programme measured only against your tracked head terms will therefore systematically under-observe the searches actually generating your exposure.
• Absence is ambiguous. Google says AI Overviews are shown only where its systems judge them additive to classic Search, and so often do not trigger at all. Not appearing in a feature that never fired is not a loss, and reporting it as one manufactures a problem to sell against.

So the engagement fixes, before any work starts, which metrics count, which report each one comes from, and which questions will be answered with a dated sample instead of a number. Naming the limits up front is the difference between a measurement commitment and a reporting habit — the limits do not become less true if they are disclosed after the invoice.

Decision three: how much of your query mix is exposed to an AI answer at all

Exposure is not uniform, and scoping without knowing your own exposure is how budgets end up in the wrong discipline. Independent evidence on the demand side comes from the Pew Research Center, which analysed the browsing of 900 US adults who agreed to share their activity: 68,879 unique Google searches from March 2025, of which 12,593 produced an AI summary, with the result pages collected between 7 and 17 April 2025.

Two caveats we state rather than bury: this is a US panel, not an Indian one, and it is a 2025 snapshot of a fast-moving surface. Treat it as a shape, not as your number.

The shape is what matters for scoping. Exposure rises sharply with query length and with question-shaped phrasing. If your demand is concentrated in two-word brand and category terms, your structural exposure to answer surfaces is low, and a plan weighted heavily toward answer formatting is over-scoped for you. If your demand is long, question-led research — comparisons, eligibility, process, cost — your exposure is high and the click economics behind it are visibly worse. That distinction is answerable from your own Search Console query export, on your own data, before anybody signs anything.

Decision four: why citation-chasing should not come before the on-page basics

Corroboration work — earning independent mentions and references so an engine has a reason to repeat what you say — is slow to pay back, because it depends on other people choosing to mention you. It is also tempting to put first, because it presents well in a pitch.

Our census supplies a base rate that argues against that ordering. Of the 1,944 reachable Wikidata-listed Indian business home pages that permit OpenAI’s search crawler, only 445 — 22.9% — met all four on-page checks we tested: JSON-LD present, an Organization entity, a self-referencing canonical and no noindex directive. More than three in four had not. None of those checks is an AI-specific requirement, and all four are within your own control, whereas corroboration depends on other people; buying the slow discipline while those basics are unmet is a sequencing error, and it is visible in a proposal before any money moves.

There is peer-reviewed support for scoping this per situation rather than from a template. The paper that introduced the term GEO — Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, “GEO: Generative Engine Optimization”, KDD 2024 — reports that the effectiveness of its optimisation strategies varies across domains, and concludes that domain-specific methods are needed. That is the academic form of the same argument: what works in one category is not transferable as a checklist to another.

One honesty note about that paper, since its headline figure is easy to misread. Its headline improvement figure is measured on GEO-bench, a research benchmark of queries and sources built by the authors. It is a benchmark result, not a client outcome, and we do not quote it as one. Anyone presenting that number as an expected uplift for your business has changed what it means.

Consultant, one-time audit, or managed team — which purchase is right

Three different purchases solve three different problems, and buying the wrong shape costs more than buying the wrong tactic. The comparison below is about purchase shape only; the one-time audit carries a published figure on the deep SEO audits page and the managed programme a published starting rate on our pricing page, while this engagement is quoted by proposal — and for the managed programme and this engagement, scope, exclusions and figure are agreed in a written proposal before anything is billed.

The broader trade-off between an individual consultant and a managed team — cost, coordination overhead, single points of failure, who covers you when one person is unavailable — is argued in full in our comparison of an SEO consultant versus a managed SEO team. It is a real decision with real trade-offs in both directions, and it is not settled by whichever page you happen to be reading.

What this engagement will not promise, and what we will not put in the deliverable

Every constraint below is a refusal we can be held to, which is the only kind worth publishing.

• No guaranteed citation or answer inclusion. No provider controls a generative engine’s source selection, and Google states that indexing and serving are not guaranteed even for a page that meets every requirement and policy.
• No proprietary tactic list presented as an AI-search method. The platform documentation says no such list is required. Writing one and charging for it would be inventing a product.
• No traffic forecast for AI features. Search Console does not isolate that traffic, so the forecast would have no denominator. We will forecast against metrics that have a source report, or not at all.
• No time-to-result date. We do not have one that is honest, and a date invented to close a deal is a date you will be measuring us against for the wrong quarter.
• No llms.txt-style file recommended as an AI-search ranking measure. Google’s guidance says no new machine-readable or AI text file is needed for eligibility. If you want one for other reasons, that is your call; it will not be listed as a visibility tactic.
• No competitor “AI share of voice” percentage. It has the same reproducibility problem as our own sampling, and a number nobody can reproduce is not evidence, however confidently it is charted.
• No crawler-permission change made on your behalf without a named owner on your side agreeing it. It carries legal and commercial consequences that are not ours to accept for you.

There are also limits that are nobody’s fault. We cannot promise that your category is researched through assistants at all — in some categories it may not be, and finding that out is a legitimate outcome of the diagnosis rather than a failure of it. We cannot promise the reporting stays as it is; Google has changed what Search Console exposes before and may again. And we cannot promise that acting on a correct diagnosis produces a commercial result, because a diagnosis governs where effort goes, not whether the market wants what you sell.

Questions buyers ask