LLM optimization agencies: what are you actually buying?

Public AI-search optimization concerns information available on your website and other public sources. An agency can improve those materials and study observed responses. It does not control a platform's model, source selection or organic recommendations.

Model training is a different process. OpenAI explicitly distinguishes GPTBot, associated with potential training use, from OAI-SearchBot, which supports its search features. Allowing a crawler is not a contract for a citation or a way to force a model to learn a particular claim.

Retrieval-augmented generation, or RAG, in your own application is different again: your team can control its document collection, retrieval pipeline and evaluation set. If this is the actual requirement, hire for application engineering rather than assuming a marketing retainer includes it.

A public-discovery engagement should identify the URLs and facts it will improve. Useful work can include making important content available in HTML, correcting canonical signals, clarifying product and entity descriptions, and linking answers to the source pages that support them.

Structured data should reflect the visible page, not add invented authority, reviews or capabilities. Google says there are no additional technical requirements for its AI search features beyond ordinary search eligibility. Treat claims that one special file or schema guarantees inclusion as a warning sign.

Ask to see the underlying records, not only the final chart. A response can mention a brand without linking to it, or cite a page without recommending its business. Count those outcomes separately and document how ambiguous examples are reviewed.

Repeated observations should use the same prompt panel and record the platform, product surface, date, locale and account conditions where known. If the provider changes the panel or model, that belongs in the report. A technical audit score must not be relabeled as a live visibility score.

Use the named agency examples below to compare published public-search services. For internal assistants, ask a prospective engineering partner about document access controls, retrieval evaluation, logging and privacy instead. A software tool that reports mentions is not automatically a substitute for either kind of implementation.

Be explicit about exclusions. Content production, software subscriptions, development, third-party outreach and manual response review may be separate workstreams. A narrow, verifiable scope is more useful than a promise to optimize every language model at once.