Key facts
- No vendor can guarantee a citation inside a ChatGPT, Gemini, or Perplexity answer, because each model decides what to surface at the moment someone asks, and the same question can return a different set of sources on different runs, even with nothing changed on your site.
- llms.txt is a proposed convention, not a Google or OpenAI standard, that lists a site's key pages in plain text for a language model to read; some crawlers respect it, none are required to, and it is a small piece of a larger program, not the program itself.
- A page written to be quoted by an AI answer engine benefits from a short, direct, self-contained answer near the top, clear author or business identity, and specific facts a model can lift cleanly, which is closer to how this pack's own family of pages is built than to a typical blog post.
- AI Overviews and chat answers frequently cite pages that also rank well in classic organic search, so a program with zero traditional SEO foundation under it is unlikely to gain AI citations either; the two are connected, not separate budgets.
- There is no equivalent of Google Search Console for most AI answer engines, so measuring citation frequency relies on running repeatable sample prompts by hand or with a tool and logging what comes back, dated, rather than a live dashboard.
Scope boundaries
Name the concrete deliverables: restructuring key pages so they carry a clear direct answer near the top, adding or fixing FAQ and article schema, publishing or maintaining an llms.txt file, and building or updating pages that answer the specific questions your buyers are typing into a chat window. Include a defined cadence for running sample prompts across the major AI engines and logging what each one names.
Out of scope, and this needs to be explicit: guaranteeing a citation in any specific AI answer, guaranteeing removal from a competitor's citation, and anything described as "AI ranking factors", since no engine publishes a ranking algorithm the way Google historically has for classic search. If a bidder's proposal implies certainty about how a model chooses sources, that is the scope to push back on.
What the vendor must be given
CMS access to edit page structure, headings, and schema markup, not just permission to write recommendations you have to implement yourself. Access to your existing analytics and search console data, since a program that ignores your current organic footing is starting from nothing when it does not need to.
A list of the actual questions your sales and support teams hear from prospects, which is often the single most useful input for this kind of work, since it tells the vendor what a buyer might type into a chat assistant. Sign-off authority on one person to approve schema and structural changes quickly, since these changes tend to be small and numerous rather than one large project.
Ownership and exit clauses
Every page edit, every schema change, and every llms.txt file the vendor writes belongs to your site outright and stays in place if you cancel, since none of this work should live outside your own CMS or codebase in the first place. There is no separate "AI account" a vendor can hold hostage the way a domain or ad account can be.
Require delivery of the full prompt log used to test citations, meaning the exact questions run and the dated results, in an editable format, so you keep the record of what was actually measured rather than losing it the day the contract ends.
Milestones and acceptance tests
A baseline prompt run across the major AI engines before work starts, logging what is currently named for a defined list of questions relevant to your business. Structural and schema changes on an agreed set of pages within the first 60 days. A second prompt run at 90 days using the exact same questions, so the comparison is honest.
Acceptance test: the 90-day report should show the baseline, the current result, the date each run happened, and a plain statement of what changed and what did not, including a citation that appeared for a competitor and not for you. A report that only shows improvements and skips what did not move is not a complete report.
Scoring rubric
Weight bidders toward honesty about limits as much as toward technical skill.
Understanding of what is and is not measurable, and willingness to say so in writing, 25 percent.
Quality of a sample page restructure using one of your real pages, 25 percent.
Schema and technical implementation ability, 20 percent.
Measurement method and reporting format, 20 percent.
Price and contract length, 10 percent.
Any bidder who guarantees a specific citation outcome should score zero on the first criterion regardless of the rest of their proposal.
Questions every bidder must answer
Ask these in writing.
What specifically will you change on our site, page by page, and how will I verify it was done?
How do you measure whether AI citations changed, and can I see a sample report from an existing client?
What is your position on guaranteeing a citation in a specific AI engine's answer?
How does this work connect to, or overlap with, our existing SEO, and are we paying twice for the same page?
What happens to the llms.txt file and schema changes if we cancel?
Which AI engines does your sample prompt testing actually cover, and how often do you re-run it?
Related questions
No. Every answer engine generates its response at the moment someone asks, weighing sources in ways none of them fully publish, and the same question can return different sources on different runs. A vendor guaranteeing a specific citation is promising something outside their control; a real program improves the odds, it does not lock in the outcome.
It overlaps heavily with SEO rather than replacing it. Pages that already rank well and are structured with a clear direct answer tend to get cited more often by AI engines, so a program built on zero existing SEO foundation is starting from a disadvantage. Treat this pack as an addition to solid SEO, not a substitute for it.
It is a proposed, informal convention, a plain text file listing a site's key pages for a language model to read, not an official standard any engine is required to follow. Some crawlers respect it. It is a reasonable, low-cost piece of a broader program, but it will not move much on its own.
Run the same set of real customer questions through the major AI engines on a schedule, log the date and the exact sources each one names, and compare over time. There is no equivalent of Search Console for most of these engines yet, so this manual, dated logging is currently the honest way to track it.
Given how new and how variable this space is, a shorter initial term with a clear 90-day acceptance test, as scoped in this pack, protects you better than a long commitment based on promises about an outcome no vendor actually controls.
SearchPod is a vendor for this kind of work and would answer this RFP; the acceptance tests here are the ones we agree to.
Want this scoped and priced for your business?
Get a free, no-obligation proposal within one business day. We look at your site and your market and tell you plainly what we would do, and what we would not.
Get your free proposal