SIGNALS
Buying AI visibility

What should you put in an RFP for AI search visibility?

The short version

An RFP for AI search visibility needs four things a normal SEO RFP does not have: which engines the vendor measures and how often, how the vendor separates a real movement from the week-to-week noise in its own data, what you keep if you leave, and what the vendor is explicitly not promising. Scope, timeline, references and commercial terms carry over from any marketing RFP unchanged. The measurement section is the one that decides whether you can ever tell whether the money worked, and it is the section most published templates leave as the single instruction to track AI visibility. Ask for the prompt list, the run frequency, the baseline design and the written definition of a real move, before you shortlist anybody.

What should you put in an RFP for AI search visibility?

Put four sections in an AI search visibility RFP that you would not put in an SEO one: an engine and prompt coverage requirement, a measurement design requirement, an exit and data portability requirement, and a statement of what you are not asking anybody to guarantee. The rest of the document, background, scope, timeline, references, pricing and scoring, is ordinary procurement and should be reused rather than rewritten.

The reason for the split is that the familiar sections of a marketing RFP all assume a scoreboard that exists. An SEO RFP can say "report monthly on target keyword positions" because positions are published by the tool everybody shares. No equivalent exists here. The engines do not publish a visibility metric, so every vendor builds its own, from its own prompt list, at its own run frequency, and two honest vendors can hand you different numbers for the same brand in the same week. If the RFP does not specify the measurement, you have not bought a programme, you have bought a vendor's house metric.

Published AEO RFP templates do exist and the good ones are useful for the structure. Conductor's template for evaluating AEO platforms is one, and it asks buyers to compare on data collection because answer engines do not publish organic visibility metrics the way search platforms do, in its published AEO tool RFP template (read 17 November 2026). Treat any of them as a skeleton. The parts worth writing yourself are the four below, because they are the parts a vendor cannot answer with a capability list.

One practical note before the sections. Send the same questions to every respondent and require the answers in the same order. The differences that matter in this category are buried in method, and method is only comparable when it is laid out in the same shape twice.

Why does an SEO RFP buy the wrong thing for AI search?

An SEO RFP buys the wrong thing because it is built around page work and reports against rank, and in AI search the page is the smaller of the two levers. The largest study to separate the two found that which pages the engine's own search retrieves explains about 32% of the variation in citations while page properties explain 2.8%, according to Mahe Chen and Nick Koudas in "What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization" (arXiv 2610.15033, submitted 10 October 2026). Their study covered 158,099 answers and 468,037 citations across two 90-day collections.

Two things follow for the document in front of you. A scope section that lists only page deliverables is buying the mechanism Chen and Koudas found that page properties explain, 2.8% of it, and calling it the programme. And a reporting section that asks for a position is asking for a number that does not exist, which pushes the vendor into inventing a proxy and reporting that instead.

The same study is worth reading for the part that is awkward for everybody selling in this category, including us. Chen and Koudas also found that the engine names brands it already knows, favours sites it recognises, and that changing only a brand name moved product rankings when the underlying facts were identical. Their own summary is that pages are cited when searched and brands are also named when remembered. An RFP written as though page edits alone will change who gets recommended is written against the evidence.

So the honest shape for the scope section is two columns rather than one: what the vendor will change on your pages, and what the vendor will do about presence off your site. A vendor who proposes only the first should say so plainly, and a vendor who proposes the second should say who does the work and what it costs. Our comparison of which kind of agency you hire to get named in AI answers covers which firms actually do each half.

What should the measurement section of the RFP require?

The measurement section should require five specific things in writing: the prompt list, the engines covered separately rather than averaged, the run frequency, the baseline design, and the vendor's written definition of a movement worth acting on. Those five are what turn a monthly report into evidence instead of a screenshot.

Require the baseline to be run more than once before any work starts, because the first thing you need measured is your own noise. BrightEdge's citation volatility analysis found that frequently cited domains moved about 0.7% week to week while sporadically cited domains swung by more than 50%, a 70-fold gap it reported as holding across ChatGPT, Perplexity, Google AI Overviews and AI Mode, and it put the inflection point near 50 citations, where weekly volatility drops from 50% to 8%, in "AI Search Engine Citation Volatility: The 70x Stability Gap" (23 October 2025, read 17 November 2026). It is one vendor's dataset with limited method disclosure, and it is still the clearest published statement of the problem: below that volume, a missing citation is one sample.

Require the run history rather than the report. Trakkr's citation decay study logged 108,650 citation URLs and 857,138 daily reports covering 10,991 brands across eight models, and found that 73.5% of citation URLs appeared exactly once and did not return, with the median brand falling to half its highest observed day in 31 days, in "The Half-Life of AI Citations" (published 22 April 2026, read 17 November 2026). A single reading describes a brand for about a month, so the series is the deliverable and a monthly PDF is not.

How large the prompt set needs to be is a separate question with its own arithmetic, and our guide on how many prompts you need to track AI visibility works through the sample size and the run count. Ask the vendor for a number and the reasoning behind it, not a tier name.

What deliverables should the RFP ask for?

Ask for deliverables that outlive the engagement, and be specific about the form each one arrives in. The useful test is whether a replacement vendor could pick the work up from what you hold without repeating the first two months, because that is exactly what you are paying for twice if you cannot.

Five items cover most engagements. The prompt set, as a file rather than a view inside somebody's platform. The full run history, with the prompt, the engine, the date and every cited URL for each run. A page change log, by URL and date, so that a movement three months out can be lined up against something real. The diagnostic that says which stage a page is failing at, rather than a score with no attached reason. And a written record of what was tried and did not work, which is the deliverable nobody asks for and the one that saves the most money later.

Be careful about the opposite failure, which is an RFP that asks for volume. Twelve articles a month is a specification a vendor can meet while the programme goes nowhere, and it creates pressure to publish pages that cannot be grounded. Our own rule on that is public and it is worth borrowing into your scope section: where a page is too thin to support a claim, the correct output is to abstain rather than to fill the space. What an engagement actually contains month to month, as opposed to what gets listed in a proposal, is set out in our guide on what an AEO engagement includes.

One more line worth adding: name who approves page changes on your side and how long approval takes. The commonest cause of a slow first quarter is not the vendor, it is a review step nobody scoped.

What should the RFP say about guarantees?

The RFP should say about guarantees that you are not asking for any: no guaranteed citations, mentions or rankings, and a response offering them scores lower rather than higher. That single sentence is the cheapest quality filter in the document, and it changes which firms respond.

The reason to put it in writing is that the thing being promised is not in the vendor's control. Retrieval is the dominant term, as the Chen and Koudas figures above show, and model memory of your brand is a second channel that no page edit reaches directly. A vendor who guarantees an outcome built on those two is either not reading the evidence or is counting on you not to. Our longer answer on this sits in whether an AEO agency should guarantee results.

What a vendor can commit to is narrower and perfectly checkable: a measurement cadence, a named set of pages brought to a stated structural standard, a diagnosis of the stage each page fails at, and a report that separates what moved from what is noise. Ask for commitments of that shape and the responses become comparable.

Ask also what evidence of past results the vendor can actually show, and accept that in this category it will often be thin and anonymous. Ours is: we publish two anonymised early case studies where one client's audit score went from 17 to 68 and another more than doubled, with the caveat stated on the page that the audit score is an input metric and share of voice is the outcome. A vendor that volunteers the difference between those two things is telling you something useful about how it will report in month six.

How should you score the responses?

Score the responses on method rather than on fluency, and weight the measurement answers highest, because every other difference between two vendors in this category is recoverable and that one is not. The table below is the scoring sheet we would use, with the answer that should worry you alongside the answer that should not.

What you asked A strong answer A weak answer
Engine coverage Each engine reported separately, with the prompt list per engine One blended visibility score across engines
Baseline design Two runs before any work, to size your own variance One reading, treated as the starting position
Definition of a real move A stated rule, such as the same direction across three consecutive runs Any week-on-week rise, shown as a chart
Prompt set sizing A number with the reasoning, and a rule for when it changes A tier name, or whatever the platform includes
Off-site work Named, scoped and priced, or explicitly out of scope Implied by the words authority and presence
Exit Prompt set, run history, change log and failures, exportable Reports available while the subscription is active
Guarantees None offered, with the reason given A citation or ranking promise

Weight the first three rows at roughly half the total. A firm that answers those three well will be answerable for the rest of the engagement, and a firm that cannot answer them will produce a report you have no way to argue with.

If you want the measurement answers in front of you before you write the document, a free visibility assessment produces them from outside your systems: your buying questions sampled across ChatGPT, Perplexity, Google AI Overviews and Claude, plus a read of your public pages as a crawler sees them. Running it first tends to shorten the RFP, because you stop asking vendors to tell you where you stand and start asking them what they would do about it.

What else do people ask about an AI search visibility RFP?

Procurement and marketing teams writing an AI search visibility RFP for the first time ask these five questions most often.

Do you need a separate RFP for AI search visibility, or can you add a section to an SEO RFP?

A section added to an SEO RFP works, as long as the section is the measurement one. The scope, reference and commercial parts of an SEO RFP carry over unchanged. What does not carry over is the assumption that the vendor reports against a ranking, because there is no ranking to report against, so a bolted-on paragraph asking vendors to describe their AI search approach gets you marketing copy rather than a method.

How many vendors should you send an AI visibility RFP to?

Three to five, and ask all of them to answer in the same format so the responses can be read side by side. The category is young enough that a wider net mostly adds firms describing the same tactics in different words, and the differences you care about show up in the measurement answers rather than in the list of deliverables.

Should the RFP ask vendors for their pricing?

Yes, and ask what changes it rather than only what it is. Scope, the number of pages in play, how much writing the vendor does rather than reviews, and how many engines and prompts the measurement covers are the four things that move a quote in this category. A figure with no scope attached cannot be compared with another figure with no scope attached.

What should the RFP say about guaranteed citations or rankings?

State plainly that you are not asking for guaranteed citations, mentions or rankings, and that a response offering them will be marked down. Removing the incentive to promise an outcome nobody controls is the cheapest quality filter in the whole document, and it changes who bothers to respond.

What should you require the vendor to hand back at the end?

The prompt set, the full run history behind every report rather than the report summaries, the page changes themselves, and a record of what was tried and did not work. Those four items are what make the next engagement cheaper than this one, and a vendor whose platform cannot export them is selling you a subscription to your own measurement.

Related guides

The Assessment

Know where you stand before you write the RFP.

A free visibility assessment samples your buying questions across ChatGPT, Perplexity, Google AI Overviews and Claude, and reads your public pages the way a crawler sees them.

Request a free visibility assessment →
SIGNALS · A BlackSig Systems company