An AI-search proposal is reviewable when you can point at a named deliverable, a stated evidence type and an acceptance test. Most proposals fail not because the work is worthless, but because the words "visibility", "citation" and "optimisation" are used without saying which of several different questions they answer. This guide gives you a reading method, section by section, that does not require buying anything first.
Before you start, write one paragraph stating what you think you are buying. If you cannot state it, the proposal cannot be judged against it. The scope builder produces a Markdown brief — business and public site, audience, one fixed intent, evidence needed, delivery questions, success criteria and constraints — that you can compare against any proposal. Background reading sits in define an AI-search scope of work and specify evidence before visibility claims.
1. Make the proposal commit to one of four questions
Appearance, citation, indexing and traffic are separate claims with separate evidence. A page can be indexed and never cited; cited and never clicked; shown as a passage and not linked. When a proposal says "we will get you into AI answers", ask which of the four it means and what record will demonstrate it.
| Question | What it actually means | Evidence to demand |
|---|---|---|
| Indexing | The URL is crawlable and eligible to be indexed and served with a snippet. | Named URLs, the crawl/index state observed, and who checked it and when. |
| Appearance | Some part of your material is surfaced in a generative answer or result. | The engine, the query used, the date, and a saved capture. Manual observation, not a score. |
| Citation | Your page is referenced as a supporting link the reader can follow. | The referring URL or passage and a dated screenshot or archive. |
| Traffic and outcomes | People arrive and do something you value. | Your own analytics: sessions, engagement, conversions, with the referrer caveat stated. |
Where a proposal promises the fourth while only describing work on the first, that is not deception — it is a missing measurement layer. Ask for it explicitly, or price the engagement as preparation only.
2. Test technical claims against the engine's own documentation
Google's position is unusually plain. Its AI features documentation states there are no additional requirements to appear in AI Overviews or AI Mode and no special optimizations necessary, and that existing SEO fundamentals remain worthwhile. To be eligible to be shown as a supporting link, a page must be indexed and eligible to be shown with a snippet, meeting standard Search technical requirements; the same page stresses that meeting requirements does not guarantee crawl, index or serving.
Google's generative-AI optimization guide explicitly declines several tactics that appear in agency decks: creating LLMS.txt or other "special" files, "chunking" content for machine consumption, rewriting copy specifically for AI systems, pursuing inauthentic "mentions", and over-focusing on structured data as an AI lever. It also states that optimising for generative AI search is still SEO, and that "AEO" and "GEO" are third-party labels.
Implication for review: a line item whose whole value is a new machine-readable file, AI text file, markup, an AI-only rewrite or a schema addition should be traced to a documented effect. If the proposal cites none, ask the vendor to cite one, or reduce that line to a normal content or technical task with a normal justification.
3. Check crawler-access work is described as layered
Crawler permissions are the part of an AI-search proposal most likely to be half-done. Google's robots.txt introduction says robots.txt primarily manages crawler traffic and is not a reliable way to keep pages out of results: a disallowed URL can still be indexed if linked elsewhere. For OpenAI, the crawler documentation lists separate, independent controls — OAI-SearchBot for ChatGPT search, GPTBot for content that may be used in training — and notes roughly 24 hours for its systems to adjust after a robots.txt change, and that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers.
OpenAI's advertiser crawler guidance is useful even for organic work because it names the failure modes a proposal should mention: a robots.txt rule alone may be insufficient, and firewalls, CDNs, bot protection, login walls, CAPTCHAs or browser challenges can separately block access, producing 403, 404, 429 or blank/challenge pages. It also warns against relying on short-term IP observations and advises validating traffic through user-agent identification, verified bot programmes, allowlists, robots.txt behaviour and provider verification.
So a credible proposal will separate: robots directives per user agent, infrastructure-level rules, and verification that the crawler actually receives the page. A proposal that promises to "unblock AI crawlers" with one robots.txt edit is incomplete.
4. Separate observation from scoring
Microsoft's description of Copilot Search in Bing says the product cites sources prominently, inline-links passages and gives access to cited sources. Its later AI visibility features announcement is careful about limits: Intents, Topics, Citation Share and Compare were in preview, labels may be broad for niche domains, and Citation Share is observational — it does not expose competitor domains, represent traffic share, or assign quality scores. Both are Microsoft's own accounts of its products, not independent verification of citation or traffic outcomes.
Use that as a benchmark for the proposal. Reporting that is honest will name the tool, the date range, the query set and the known limits. Reporting that produces a single "AI visibility score" with no described method is not evidence, and a proposal that promises a ranking position inside an AI answer is promising something the cited vendor documentation does not support.
Google directly supports scepticism here: its guide advises wariness of third-party tools that promise ranking success or claim to use "internal" Google metrics, noting no third-party tool has access to Google's internal ranking or AI systems.
5. Convert every deliverable into a noun someone can accept
A proposal passes review when each line has: an output, an owner, a date and an acceptance test. Rewrite verbs into nouns as you read.
- "Improve AI visibility" becomes "a written page-level review of 12 named URLs against the fundamentals Google lists, with an issue list and priority".
- "Get cited" becomes "a dated observation log for a fixed set of 20 questions across two engines, recorded manually, with captures and a plain statement that inclusion cannot be guaranteed".
- "Fix crawl access" becomes "a per-user-agent access matrix covering robots directives, CDN and firewall rules, with recorded request outcomes before and after".
- "Optimise content" becomes "a named set of sections rewritten by a named person, with sources and an authorship statement".
Google's people-first content guidance is a good quality gate for content-shaped deliverables: it asks for originality, demonstrated effort, clear sourcing, accurate authorship and AI/automation disclosure where a reader would reasonably ask how something was made, and it warns that fabricated author profiles or credentials are deceptive. Any proposal offering to "add author authority" should say whose real, verifiable expertise is being described.
6. Five questions that decide the review
- Which of the four questions does each deliverable answer, and what record proves it?
- Which engine and which surface — Google AI features, ChatGPT search, Copilot Search — and what does that vendor's own documentation say?
- Is measurement manual observation, first-party reporting, or neither? If a tool, which one, and what are its stated limits?
- Who does the work, what are they accountable for, and what happens if an engine changes behaviour mid-engagement?
- What is deliberately excluded — and does the exclusion match your constraints and budget?
If the proposal survives these five, the remaining negotiation is commercial. If it does not, no amount of scope expansion will fix it. Compare notes against citations, referrals and outcomes, then draft the counter-scope in the scope builder and send that back as the basis for the next version.
Methods and limits
This guide synthesises vendor documentation and a reviewing method; it contains no client results, no team claims, no ranking estimates and no automated third-party query checks. Statements about Google and OpenAI describe those companies' own systems as documented, not other engines. Documentation pages change; check the dates on the cited pages before relying on a specific detail in a contract. Microsoft's cited posts describe its own products and features that were in preview at publication.