A third of B2B software companies now publish an llms.txt file. In my crawl of 3,877 sites, 1,041 of 3,233 have one. Ahrefs found 97% of those files received zero fetches in a month.
Somebody sold most of those files. Somebody billed for them.
That is the market you are shopping in. Every proposal promises AI visibility. All three look identical until you ask two questions nobody puts in a deck.
What does an AEO consultant actually do?
Three different jobs, usually sold as one.
Measurement. Running prompts across engines, recording what gets named, and keeping the protocol stable enough that the number means something in 90 days.
Technical work. Crawler access, server-rendered claims, crawl hygiene, entity markup. This belongs to whoever controls the site and the CDN, which is rarely the person signing the contract.
Positioning. Deciding what you say that a competitor could not say. This is the one that cannot be delegated to a contractor, and the one most proposals skip.
Ask which of the three a proposal covers. A retainer that quietly contains only the second one is an SEO contract with new vocabulary on the cover.
Ten questions to ask, and the answers you want
Work through these in the first call. They sort the market in about fifteen minutes.
- How many prompts, how many runs per prompt, which engines? You want a fixed set, five or more runs, and three or more engines. One run per prompt is a coin flip recorded as a fact.
- Logged in or logged out? Logged out. A signed-in session puts personalization into your report.
- Do you record the model version on every observation? Yes, or movement is unattributable.
- What do you show me when the whole platform moves? On October 18, 2025, Profound measured an average 31% visibility drop across tracked brands from one ChatGPT change. A provider with no platform baseline will sell you that day as your own decline.
- How many sections do you score, and does the list change between clients? Fixed list, or the percentage compares to nothing.
- Can I reproduce your number myself? If the method is private, the number is unauditable.
- Do I get the raw observations and a baseline file I keep? Yes, and you keep them when the engagement ends.
- Who owns each fix? Named roles, not “the team.” Most technical fixes belong to someone who was not on this call.
- What will you refuse to recommend? See the next section.
- What would make you tell me not to hire you? A provider who cannot answer this has never done it.
The refusal test
Ask what they will not recommend. This single question separates the market faster than any case study.
Five tactics have been tested and failed. A provider should be able to name most of them without prompting:
- llms.txt. SE Ranking found no measurable relationship to citations across about 300,000 domains.
- Schema volume for citations. Ahrefs added schema to 1,885 pages and AI Overview citations fell 4.6%. Organization markup still earns its place for entity disambiguation.
- Blocking Google-Extended to exit AI Overviews. It is a training opt-out token, not a crawler.
- Review volume as a lever. A profile is an eligibility gate. Seer found 1% citation rates with no profile against 53.5% with as few as 1 to 13 reviews. Volume past that explains under 2% of variance.
- Chasing Bing rankings. Bing indexation is a gate worth clearing in Webmaster Tools. Bing position is not a visibility KPI.
A review profile is a gate, not a volume lever
AI citation rate for brands with no review profile against brands with as few as 1 to 13 reviews.
Brands with as few as 1 to 13 reviews were cited at 53.5%, against 1% with no profile. Review volume past that explains under 2% of variance, so a provider selling review volume as an AI visibility lever is selling the wrong thing.
Sources: Seer (citation rates by review presence). G2 (share of variance explained by review volume).
The full reasoning behind each refusal is here. A provider with no refusals is selling a checklist, and a checklist is cheap to produce and expensive to work through.
Can you reproduce their number?
If not, you cannot defend it internally, and you cannot tell progress from weather.
Ask for the prompt set in writing. Ask for the run count. Then run three of those prompts yourself, logged out, five times each, and see whether your counts land in the same range. The method takes an afternoon.
A provider who hands over the prompts is confident. A provider who calls the prompt set proprietary is protecting a number rather than a method.
Pricing shapes, and what each one hides
Three shapes, and each one carries a different risk to you.
Audit first, fixed fee. You buy a diagnosis, see the quality of thinking, and decide afterward. The risk is a report that sits in a drawer, which is why the deliverable should include a fix list with owners rather than observations alone.
Monthly retainer. Predictable for them, open-ended for you. Reasonable once the work is known, expensive while it is still being discovered. Ask what happens in month four if the number has not moved.
Project fee for a defined scope. Clean, provided somebody has already diagnosed the problem. Skipping the diagnosis is how a client pays for schema work when the actual failure was a firewall rule.
Whatever the shape, ask what you keep. The prompt set, the baseline file and the raw observations should be yours.
Who does not need to hire anyone
Say it plainly: plenty of companies do not.
Skip the hire if four things are true. You rank well. You publish original research. You already get named in answers across several engines. And a developer on your team can read a server response. Run the count before you spend.
Hire someone if your category question returns four competitors and not you. Or if nobody can tell you whether your homepage claim sits in the server HTML. Or if your claims would survive a competitor copying them word for word.
The buyer path matters here too. Gartner surveyed 645 B2B buyers between August and September 2025. 45% said they used generative AI during a recent purchase. Separately, 69% said they prefer to validate AI-generated insights with a sales rep. The answer builds the shortlist. The rep confirms it.
Many B2B buyers use AI to buy. Most want a rep to check it.
Two separate questions from one survey of 645 B2B buyers. Each bar shows a share of all 645 buyers.
The answer builds the shortlist. The rep confirms it.
Source: Gartner survey of 645 B2B buyers, August to September 2025.
Five red flags
Each one has cost somebody a quarter.
- A visibility score reported to two decimal places. The underlying data does not support that precision.
- A guaranteed number of citations. Nobody controls a retrieval system’s output.
- A deliverable list with no diagnosis in front of it. That is a guess with an invoice attached.
- Percentages with no denominator. “560% growth in 45 days” means nothing without the base.
- No named owner for technical fixes. The engagement stalls in week two when nobody can change robots.txt.
Ask this before you sign
One question, and it reveals everything the deck does not. Which of my four conditions am I failing, and what evidence says so?
The four are access, readability, a claim worth selecting, and presence on the sites the engines already cite. A provider who answers with deliverables instead of a diagnosis has not looked at your site.
My own answer is a fixed-fee diagnostic. It runs 150 observations across ChatGPT, Google AI Mode and Perplexity. It scores 30 sections that do not change between clients. It tests your claims against quoted competitor sentences. You keep the baseline file. What belongs in an audit is here, including the questions to ask me.
Frequently asked questions
What does an AEO consultant actually do?
Three different jobs, usually sold as one. Measurement, technical work, and positioning. Ask which of the three a proposal covers. A retainer that quietly contains only the technical work is an SEO contract with new vocabulary on the cover.
What should I ask in the first call?
The ten questions sort the market in about fifteen minutes. Ask how many prompts, how many runs per prompt, and which engines. You want a fixed set, five or more runs, and three or more engines. Ask whether they run logged out, record the model version on every observation, and let you reproduce the number yourself. Ask who owns each fix, what they will refuse to recommend, and what would make them tell you not to hire them.
What is the refusal test?
Ask what they will not recommend. This single question separates the market faster than any case study. Five tactics have been tested and failed. A provider should be able to name most of them without prompting: llms.txt, schema volume for citations, blocking Google-Extended to exit AI Overviews, review volume as a lever, and chasing Bing rankings. A provider with no refusals is selling a checklist.
How can I tell if their visibility number is real?
Ask whether you can reproduce it yourself. If the method is private, the number is unauditable. Ask for the prompt set in writing and the run count. Then run three of those prompts yourself, logged out, five times each, and see whether your counts land in the same range. A provider who calls the prompt set proprietary is protecting a number rather than a method.
Who does not need to hire an AEO consultant?
Plenty of companies do not. Skip the hire if four things are true. You rank well. You publish original research. You already get named in answers across several engines. And a developer on your team can read a server response. Run the count before you spend.
What are the red flags?
A visibility score reported to two decimal places. The underlying data does not support that precision. A guaranteed number of citations. Nobody controls a retrieval system’s output. A deliverable list with no diagnosis in front of it. Percentages with no denominator. And no named owner for technical fixes.
Scott King helps B2B technology companies build AI marketing systems and create differentiated content.