Buying AI Search Optimization: What Should the Supplier Prove?

Before funding AI search optimization, ask the supplier to name the system they want to influence, the change they will make and the evidence that would justify further spending. A delivered file or an improved audit score can confirm that work was completed. It takes separate evidence to show better search visibility, useful visits or qualified enquiries.

This distinction matters when a proposal includes llms.txt, special AI markup or a recurring package described as GEO, meaning generative engine optimization. Each proposed change needs a purpose you can assess.

Start with the system and the intended result

Google Search and an agent consulting public developer documentation are different targets. A change intended for one needs evidence relevant to that use.

Google’s generative AI optimization guide, updated on 10 July 2026, says llms.txt does not affect visibility or ranking in Google Search. Its guidance covers Google Search’s generative features. Google also distinguishes eligibility from guaranteed indexing or serving. A supplier proposing llms.txt as a Google ranking requirement should revise that claim and explain which other proposed work has a supported purpose.

The file can have a separate intended use. Jeremy Howard’s llms.txt proposal describes guidance and linked documentation for agents. The current v2 proposal was modified on 10 August 2026. It provides a possible basis for an agent documentation experiment; it does not establish a compulsory provider policy or demonstrate search uplift.

Understand what the audit actually checks

A tool score is useful when you know its rules. Chrome Lighthouse’s llms.txt audit, updated on 5 May 2026, concerns agentic browsing. It flags server errors. A missing file returning 404 receives N/A because the file is optional.

That check can identify a technical problem with an existing file. It does not show that prospective customers’ agents use the file, or that adding one improves Google rankings. Ask the supplier to connect the finding to the named system and customer task before treating it as a reason for further work.

Use research within its limits

Ahrefs’ study published on 15 June 2026 examined May requests across 137,210 domains using Ahrefs Web Analytics. It reported limited fetching of llms.txt files. This is useful evidence for questioning assumptions about adoption, with several limits: the customer population skews technical and SEO-aware, requests were classified using responses, channels and user agents, and files were screened for error content rather than explicitly validated against the specification.

The observation period preceded the August v2 revision. It cannot evaluate that revision or establish October behavior. The study also did not measure causal improvements in citations or conversions. Its request data gives a bounded observation, rather than a forecast for your business.

Ask for a short proposal record

Before approving the work, get these points recorded in a form you can return to at review:

• Target and task: Which search product or agent, which audience and which task should benefit?
• Deliverable and ownership: What pages, files or settings will change? What will be handed over, and who will keep it accurate? Separate repair of an identified problem from experimental work.
• Evidence: Which current provider guidance or reproducible observation connects this change to the target? Include the source and any uncertainty.
• Baseline and observation: Record the present state, the useful change to look for, the available measurement and a review date. State any gaps in access or reporting.
• Review decision: What would support continuing, changing or stopping the work? For recurring fees, identify the next action that the observations would justify.

Keep the measurements distinct. A fetch does not establish that an agent read or used the contents, trained on them, cited your business, produced a human visit or generated an enquiry. A citation may produce no click; a visit may produce no enquiry. Unavailable reporting means the result is unknown, rather than zero. Even an observed improvement after delivery needs care before attributing it to that change.

Separate completion from continued funding

Consider two hypothetical proposals. An ordinary business is offered llms.txt as a Google Search ranking requirement. The buyer can request a corrected proposal against Google’s guidance, then assess any supported repairs or other work on their own merits. Installing the file would establish completion of that deliverable, without validating the ranking claim.

A business with public developer documentation has a different question: can a named agent find and correctly use current instructions for a specific task? A small optional experiment could compare the present behavior with behavior after the documentation change. Agree the task, maintenance owner, review point and stop condition before expanding it. Continue only if the observations support useful work within that scope, while acknowledging uncertainty about cause.

Accept a repair when the agreed defect is resolved. Decide separately whether the evidence supports another experiment or a recurring commitment. If you want to discuss your audience, existing site and proposed work, our services overview is a starting point.