Most AI-search disagreements are not arguments about strategy. They are arguments about what counts as proof. A brief that names its evidence in advance turns a vague promise — "we'll get you into AI answers" — into a checkable claim you can accept, reject or defer.
This guide sets out a defensible evidence specification you can write yourself, before commissioning any work. It uses Google's, OpenAI's and Microsoft's own published documentation as the reference points, because those are the vendors for which this publication holds primary sources. Where something is unknown or unmeasurable without vendor tooling, the brief should say so.
Why the evidence must come before the claim
Google states plainly that there are no additional requirements to appear in AI Overviews or AI Mode and no special optimizations are necessary, and that existing SEO fundamentals remain relevant (Google Search Central, AI features and your website). It also states that a page must be indexed and eligible to be shown in Google Search with a snippet to be eligible as a supporting link — and that meeting every requirement does not mean Google will crawl, index or serve the content.
That sentence is the whole reason this article exists. Eligibility is a gate, not an outcome. Any supplier who converts an eligibility criterion into a forecast has moved from documentation to speculation. Google's own guidance warns to be wary of third-party tools that promise ranking success or claim to use "internal" Google metrics, noting no third-party tool has access to its internal ranking or AI systems (Google's guide to optimizing for generative AI features).
So the brief's job is not to predict. It is to define, in writing, which observable facts would let you conclude the work happened.
The evidence ladder: four rungs, cheapest first
Write your success criteria as a ladder from evidence you can gather yourself to evidence that requires access you may not have. Each rung should be listed separately, with the owner and the artefact that proves it.
Rung one — access and eligibility, verified in-house
This is the cheapest and most neglected rung. It asks: can the relevant crawlers reach the page at all, and does the page carry the standard signals needed for Search?
For Google, the guidance is to keep the existing SEO fundamentals; Google's examples include allowing crawling in robots.txt and by any CDN or hosting infrastructure, internal links, good page experience, important content in textual form, and structured data that matches visible text (Google, AI features and your website). Google states these are examples of existing SEO fundamentals that remain worthwhile, and adds that you do not need new machine-readable files, AI text files or special markup, and no special schema.org structured data is required, to appear in these AI features.
For OpenAI, the controls are separate per user agent, and the distinction matters. OpenAI documents OAI-SearchBot as the crawler used to surface websites in ChatGPT's search features, GPTBot as the crawler for content that may be used in training foundation models, OAI-AdsBot for validating pages submitted as ads, and ChatGPT-User for user-initiated actions (Overview of OpenAI Crawlers). OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers though they may still appear as navigational links, recommends allowing OAI-SearchBot, and says it can take roughly 24 hours after a robots.txt update for its systems to adjust. It also states that ChatGPT-User is not used for automatic crawling and is not used to determine whether content may appear in search — so a ChatGPT-User entry in your logs is not evidence of search inclusion either way.
Two cautions belong in the brief. First, robots.txt is not an indexing control: Google's robots.txt introduction explains that a disallowed URL can still be indexed if linked elsewhere, and that noindex or password protection is the correct tool for keeping a page out (Google, Introduction to robots.txt). Second, robots.txt alone may not be enough to allow access: OpenAI's advertiser guidance notes that firewalls, CDNs, bot-protection services, login walls, CAPTCHAs or browser challenges can separately block crawler requests, producing 403, 404, 429, blank or challenge responses (OpenAI Help Center, Advertiser guidance for allowing OpenAI web crawlers). That article is specific to ads landing pages; we treat its checklist — check the exact URL, follow redirects, check every host, then check what the crawler actually received — as a useful diagnostic pattern to adapt to an organic access audit, not as OpenAI guidance on organic inclusion.
Rung two — the page is recorded as eligible
Record, on a dated sheet, the specific URL, the robots directive that applies to it, whether the page returns a full response to a verified crawler, and whether the content is present in textual form rather than only in script or an image. Keep the observation as a snapshot with a timestamp. If a supplier cannot produce this, the rest of their evidence is unanchored.
Rung three — presence and citation, observed not inferred
Here the honest position is narrower than most proposals admit. Google describes its generative features as relying on retrieval-augmented generation over the Search index plus "query fan-out", and says AI Overviews only trigger when its systems judge them additive — they often do not trigger at all (Google, AI features and your website). Microsoft describes Copilot Search in Bing as blending traditional and generative search and citing sources prominently with inline links and access to cited sources (Bing blog, Introducing Copilot Search in Bing).
Microsoft's Bing Webmaster Tools reporting is described by Microsoft itself as observational: Citation Share is calculated as the percentage of citations attributed to a site out of all citations for the same grounding query, and Microsoft states it is not a ranking system, does not expose competitor domains, does not represent traffic share and does not assign quality scores (Bing blog, New AI visibility insights). Those features were in preview, and Microsoft warns labels may be broad for niche domains.
The design consequence: if a proposal promises increased citation share and your only route to that number is a vendor dashboard, the claim is not independently checkable. Write that limitation into the brief rather than letting it be discovered later. This publication does not use Google Search Console in its own method; if a supplier's evidence depends on data you cannot see, treat that as a disclosure requirement, not a deliverable.
Rung four — outcome: traffic and downstream behaviour
Only here do you reach the question the business actually cares about. The brief should define the traffic measure in terms you control — sessions from the engines in scope, landing-page visits, enquiries — with a baseline period, a comparison window and a stated tolerance for seasonality. Google notes that clicks from results pages with AI Overviews were, in its observation, higher quality in the sense of longer time on site (Google, AI features and your website); that is vendor observation, not a planning assumption, and should not be written into a forecast.
What the brief should explicitly refuse
A strong evidence section is as much a list of exclusions as inclusions. State that you will not accept:
- ranking or citation estimates presented as measurements;
- claims of access to a search engine's internal systems;
- "mentions" or coverage counts from undisclosed sources;
- content-volume targets used as a proxy for quality.
Google's mythbusting section supports several of these: it advises against "chunking" content for AI systems, rewriting content specifically for AI, keyword-variation padding, seeking inauthentic "mentions", and over-focusing on structured data as an AI-search tactic (Google, generative AI optimization guide). Separately, Google's people-first content guidance discourages mass-producing content across topics, extensive uncurated automation, writing to a target word count, and changing dates to appear fresh (Google, Creating helpful, reliable, people-first content). Each of those is a common line item in AI-search proposals; each should be struck in the brief before signatures.
Delivery questions that force specificity
Attach these to the evidence section. They are answerable without purchase and their answers are the real test of a proposal.
- Which exact URLs are in scope, and what is each page's current access state, recorded with date and evidence?
- Which user agents do you expect to allow or disallow, and on whose authority — the vendor's own crawler documentation?
- What artefacts will you hand over, in what format, and can we keep them after the engagement?
- For each success criterion, who holds the measurement, and can we reproduce it without your tooling?
- What will you do if the content is indexable and standards-compliant but appears in no AI feature for six months?
- Which parts of your method rest on vendor documentation and which on your own testing? Please separate them.
Question six is the one that sorts proposals. Google's guide is candid that creating unique, non-commodity, people-first content with a distinct point of view is likely to matter more in the long run than any other suggestion in the guide (Google, generative AI optimization guide), and its people-first guidance frames quality around originality, demonstrated effort, expertise, clear sourcing and accurate authorship, warning that fabricated profiles or credentials are deceptive (Google, Creating helpful, reliable, people-first content). A supplier whose evidence plan is only measurement, with nothing about how original material gets produced and attributed, has answered half the question.
Write it down before the first invoice
Assemble the rungs, exclusions and questions into a single dated document attached to the brief. Then run it through the scope builder, which turns your business, public site, audience, intent choice, evidence, delivery questions, success criteria and constraints into a private Markdown scope you control. Read Define an AI-search scope of work first for the brief's structure, and Distinguish citations, referrals and outcomes for the measurement vocabulary this section depends on. Review an AI-search proposal applies these tests line by line.
Sources: Google Search Central, AI features and your website (last updated 2025-12-10); Google Search Central, Optimizing your website for generative AI features; Google Search Central, Introduction to robots.txt (last updated 2025-12-10); Google Search Central, Creating helpful, reliable, people-first content (last updated 2026-10-05); OpenAI, Overview of OpenAI Crawlers; OpenAI Help Center, Advertiser guidance for allowing OpenAI web crawlers; Bing blog, Introducing Copilot Search in Bing (April 2025) and New AI visibility insights in Bing Webmaster Tools (June 2026). Vendor documentation describes each vendor's own systems only; Microsoft's reporting features were in preview and are described by Microsoft as observational. No forecasts, rankings or outcomes are asserted here.