Chapter 22
Search and AI answers
Search optimisation and answer optimisation are one job: be named, be recommended, and be right.
What changed
Search is usually run as a ranking business. An agency picks keywords, reports positions each month and counts the traffic. Lately a second agency, or a second line on the invoice, offers the same for AI answers under a new acronym.
The work is the same. Google says so: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary" (Google Search Central). The entry ticket is ordinary search.
What has changed is the test. Clicks have stopped being the score. Ahrefs studied 300,000 keywords. It estimated an AI Overview cut the top page's click-through rate by 58%, and calls the result a correlation (Ahrefs). And one buyer question now becomes many: Google says its AI features may use "query fan-out", issuing related searches across subtopics. The page that wins answers the follow-ups too.
What it is for
The job has three tests. When a buyer asks without naming you, you're named: that's Found. You're the one recommended: that's Preferred. And what's said about you is right, your price, terms and claims repeated accurately and backed by others: that's Believed.
It also serves Answered in part, because both kinds of answer draw on pages that settle the question.
Named isn't recommended. Picture five names in an answer, with a rival called "best for teams like yours". That's a loss that looks like a win. Most AI-visibility reports count the mention and ignore the verdict.
This build doesn't create Known. A large recent preprint found Google search interest was the strongest predictor of how often AI named a brand (Chapter 13). That interest is made by memory work. This build harvests memory; it rarely creates it. Nor does it close the sale at Bought. Where the buyer lands is the website's job.
The anatomy of a good one
A good programme is a loop. It starts with what buyers ask, ends with what the machines said, then fixes the sources behind the answer.
The buying questions, per moment (Found). Start from your map of moments and the question bank in Chapter 14, not a keyword tool. Write the questions in buyers' own long, specific words.
Answer pages, in plain text (Answered). Each buying question gets one owning page, with the follow-ups the engine will ask on the same page. Keep it in text, not in an image or a gated PDF.
Technical access (Found). Crawlable, fast and consistent. It's a gate, not a lever. Cloudflare, which says it handles traffic for 20% of the web, began blocking AI crawlers by default in July 2025 (Cloudflare). AI firms run separate crawlers for training and for search, and blocking one doesn't block the other.
Other people's pages (Believed). Citations, reviews and lists. The machine weighs what others say about you above what you say. Each source is small and the list shifts, so rebuild it each quarter from your own questions.
Sampled presence (Found · Preferred). Ask the same unbranded questions many times on the assistants your buyers use. Record named, recommended and right, and which sources each answer cites.
Fix the source, then ask again (Believed). Never chase the answer. Fix the page it came from: correct it if it's yours, earn a place or a correction if it's someone else's.
The mistake most teams make
They buy tricks. There's no special markup: Google says you don't need "new machine readable files, AI text files, or markup". Manipulating answers is now spam. Google's spam policies cover "attempting to manipulate generative AI responses in Google Search" (Google).
Hidden instructions are treated as an attack. Microsoft's security researchers found more than 50 such prompts from 31 companies over 60 days, hidden in "Summarize with AI" buttons (Microsoft). Microsoft calls it "AI Recommendation Poisoning".
The quieter mistake is buying search and AI separately. Same questions, pages and sources. Two agencies means two reports and nobody owning the gap. And most teams keep publishing on their own site while the verdict is written on other people's pages.
How you know it's working
Report three rates per moment, per assistant, with their margins: named, recommended and right. Then ask new enquiries how they found you, and log the moment. Ask buyers, because Google counts AI Overview and AI Mode traffic inside Search Console's ordinary "Web" figures (Google). Keep rankings and traffic as diagnostics, never goals.
Fieldline's search-and-answer programme
Fieldline, the invented scheduling-software firm, had a search agency and a separate AI-visibility retainer. Both reported green, because both tracked the operations question about drive time.
Then a deal stalled. The buyer's IT manager had asked an assistant which scheduling tools connect to their accounting software. Two rivals were named. Fieldline wasn't.
The team mapped the sources behind her question. Its integration guides were PDFs behind a sign-up form. A security review had blocked "all AI bots" at the CDN, including the search crawlers. And a review site's listing showed an integration Fieldline had retired.
The fixes were dull: publish the guides as text, allow the search crawlers, correct the listing. Two months later the IT question named Fieldline in about a third of answers. The stalled deal didn't wait that long; it went to a rival.
The full playbook. The Search and AI Answers Playbook, with the search-and-answer audit scorecard, the page map and the citation source map, is free at rocmode.com/playbooks.
Whose evidence this is. Ahrefs sells SEO and AI-visibility tools. Microsoft sells security products. Google, Microsoft and Cloudflare describe their own systems. No peer-reviewed source is used in this chapter beyond the preprint in Chapter 13.