The New Science of Search: How AI Agents Choose Who Gets Found

Oct 25, 2026

Search is no longer a list of ten links that people click. In 2026 it is a chain of AI systems that rewrite the customer's question, read a handful of sources, pick two or three businesses, and increasingly call or book one of them. Google, OpenAI and Anthropic each describe pieces of this in their own documentation. Put together, those pieces describe a new discipline.

The big brands are spending billions trying to win this with volume: more content, more links, more ad spend. Most of that money goes to the parts of the system that matter least. This post is about the parts that matter most: what these AI agents actually read, how they decide who to name, and what a business has to have in place for an agent to act on it.

The short version: you no longer rank pages. You get chosen by machines. Getting chosen depends on three things, and only one of them lives on your own website.

How the machines actually decide

Every major AI assistant runs the same four-stage pipeline, and each stage has its own rules for who survives it.

Four gates every business must pass to get chosen by AI

1. Rewrite  ->  2. Retrieve  ->  3. Select  ->  4. Act
Splits the      Searches each    Names businesses   Calls, texts or
question into   one; reads top   many sources       books the winner
sub-questions   results only     agree on
Win: answer     Win: rank for    Win: be talked     Win: answer
each one        each             about              and quote

A business that fails any one gate never reaches the next; most marketing spend only targets the second.

  1. Rewrite. The AI never searches the customer's words as typed. Google says AI Mode splits a question into subtopics and searches each one at the same time across multiple sources (eSEOspace, quoting Google's help page). OpenAI documents the same move: ChatGPT estimates a location from the user's IP address and turns "good restaurants near me" into a city-specific query (Cheers, citing OpenAI's Help Center).
  2. Retrieve. Each rewritten sub-question goes to a search index: Google's own for Gemini and AI Mode, Bing and third-party listing providers for ChatGPT, and a separate independent index for Claude. Only the top results come back. A business that doesn't rank for the sub-question is never seen.
  3. Select. The model reads those results and names the businesses that several independent sources agree on. In one study of local recommendations, ChatGPT named just 1.2% of 350,000 locations, while Google's local 3-pack showed 35.9% (SOCi data via Shift In Life).
  4. Act. Agents now contact the chosen businesses directly. Google's AI calls local businesses to ask about availability and pricing on the customer's behalf (Google). Personal agents like Instinct text, call and book for their users.

The insight most of the industry misses: you are not optimizing for the customer's query. You are optimizing for the sub-questions the AI generates from it, the sources it trusts to answer them, and the moment an agent tries to act.

What the agents actually read

The uncomfortable truth for website owners: the AI mostly isn't reading your website. One analysis found that only about 4% of the sources an AI answer links to are the business's own site; for local questions, the sources are usually a few "best of" listicles, the big review sites and the Google Business Profile (PlusPoint).

The sources that carry the most weight, in order of leverage:

  • Third-party review and directory sites. In one field study, ChatGPT's local picks cited HomeAdvisor, the Better Business Bureau and Angi, and never the business's own website (DEV Community).
  • YouTube and Reddit. Across a 30-million-source analysis, Reddit was the most-cited domain in AI answers, with YouTube and LinkedIn close behind (Search Engine Land).
  • Brand mentions, linked or not. Ahrefs studied 75,000 brands and found web mentions correlated with AI Overview visibility at 0.664, against 0.218 for backlinks (Ahrefs). YouTube mentions correlated even higher, at about 0.737 (Medium summary of the study).
  • Review volume and content. A study of 10,000 local businesses found review count was the strongest predictor of AI visibility, with visible businesses averaging 133 reviews against 11 (Search Counsel, citing Insites).
  • Structured business data. The Google Business Profile, Bing Places and other listing feeds supply the names, hours, services and prices agents work from.

The pattern underneath: AI systems trust what other people say about you far more than what you say about yourself. Links built authority in the old system. Being talked about, consistently and specifically, builds it in the new one.

Google: the agent that already holds the map

Google's advantage is that it owns both the index and the local data. AI Mode reached a billion monthly users within a year of launch, with query volume more than doubling every quarter (Resultsense, reporting Google I/O 2026). Four things matter most.

AI answers absorb informational clicks, not transactional ones. When an AI Overview appears, click-through rates drop by close to 60% (Search Engine Land). But Seer Interactive found transactional queries trigger an Overview only about 5% of the time, against 86% for question-format queries (Cognizo, citing Seer). Local intent is even more protected: Whitespark found AI Overviews on 15% of local queries and the map pack on 93% (Entrepreneur).

The Business Profile is now an AI data feed. Google is rendering local listings as AI Overviews with a button into AI Mode chat (Search Engine Roundtable). Gemini drafts answers to customer questions on profiles, held for the owner to approve (Fog Digital). Whatever the profile says is what the AI repeats.

Google's agent calls and books. At I/O 2026, agentic booking expanded to home repair, beauty and pet care, and the AI only contacts businesses already showing in matching local results (Digital Media Cube). Businesses without a booking link connected to their profile are skipped in the automated flow (Social Places).

Google says the GEO tricks don't work. Its own guide tells site owners they don't need llms.txt files or special markup, and points them to their local business and ecommerce details instead (Search Engine Land). For the first time, Search Console also reports how often pages appear in AI Overviews and AI Mode (Medium).

ChatGPT: the answer you can't buy

ChatGPT is smaller than Google as a source of clicks but punches far above its weight on intent. In a 100-site, 12-month dataset, ChatGPT sent 0.62% as many sessions as Google organic, yet its visitors converted at 13.8% against 3.02% (Dalin de Graff, analyzing Marie Haynes' data).

It depends on other people's data for local. OpenAI says ChatGPT may share a user's approximate location with trusted third-party providers that supply nearby business listings, and it doesn't name those providers (Cheers, citing OpenAI). It now books restaurant tables through Yelp, OpenTable and Resy (Android Headlines), and OpenAI moved purchases out of ChatGPT and into partner apps (Skift). For a local business, visibility in ChatGPT runs through the platforms ChatGPT plugs into.

Ads sit beside the answer, never inside it. OpenAI says advertisers have no ability to shape, rank or alter ChatGPT's responses (OpenAI Help). Its ad system is a separate AI model that infers what a person wants commercially and auctions that, rather than matching keywords (Forbes). OpenAI says more than 100,000 advertisers have used it (CleverBoss).

Two consequences follow. First, the recommendation inside the answer can only be earned, which makes it the scarcest real estate in marketing. Second, the ads fire during research, not at the point of purchase: in a sample of 7,318 ads, "near me" prompts triggered only 1.3%, while comparisons and "best" questions triggered far more (Sprites). ChatGPT ads reach people while they're still deciding what to buy, before they look for who sells it.

Claude: the agent that reads only what ranks

Anthropic runs three crawlers, and they do different jobs. ClaudeBot collects content for training. Claude-User fetches a page when someone asks Claude a question, and blocking it stops Claude retrieving that site for users. Claude-SearchBot indexes content to improve Claude's search results (Claude Help Center). Blocking the training crawler doesn't stop Claude citing a site; blocking the other two does.

Anthropic hasn't officially named the search engine behind Claude's web search. Brave is reported as the provider based on Anthropic's subprocessor list, but that's unconfirmed (FixAEO).

From how Claude actually works in conversation, three practical points:

  • Claude sees only the top results for each search, about ten per query with snippets. It has no maps product, business database or review feed of its own, so it leans even harder on third-party pages than Google's AI or ChatGPT.
  • Claude favors specific, citable facts. Pages with dated, concrete numbers get quoted. Vague marketing copy gets skipped.
  • Claude can act, not just answer. In the Claude apps it can use a browser or a person's computer to visit a site, compare options and fill in a form. A business whose prices, services and booking are in plain text is easy for it to use. One hiding everything behind "call for a quote" is not.

Personal agents and Grok: the customer is becoming a machine

Personal agents are the next wave. Instinct is an AI assistant people text to get things done; it uses its own phone number and computer to carry out tasks (TechCrunch), including restaurant reservations and negotiating bills (PYMNTS). Its founder says it is nearing $1 billion in annual transaction volume while still invite-only (Fortune), and Meta launched a direct rival, Muse. Neither has said how it chooses which business to contact, so the safe assumption is that it searches the web like everyone else and lands on the same sources.

The implication is bigger than any one product: more and more, the first contact a business gets is from an agent, not a person. The business that answers instantly, quotes a price and offers a bookable slot wins. Voicemail loses.

Grok is a special case. xAI says Grok searches public X posts as well as the web (MetaGrok), and one academic analysis found X was its largest single source (ContextBolt). For local businesses it matters little: one tracking firm says it has no local business data integration (GEO Scout). It's worth effort only where a brand's customers actually talk on X.

The strategy: seven plays to get chosen

The plays follow the pipeline. Each one wins a stage where most competitors aren't even competing.

  1. Map the fan-out, not the keyword. For each service, write down the sub-questions an AI will generate: price, licensing, warranty, speed, reviews, service area, emergencies. Run the customer's real questions through Google AI Mode, ChatGPT and Claude, and note what each one searches and cites. Answer every sub-question in plain text, on your site and on the sources the AI reads.
  2. Own the sources, don't just appear in them. The AI reads third-party "best of" pages, local directories and YouTube. Build or acquire those assets in your niche: a genuinely useful local guide, a trade directory, a YouTube channel that reviews and explains. Whoever owns the source shapes the answer.
  3. Engineer what reviews say, not just how many. Review count is the strongest predictor of AI visibility, and the words in reviews are what AI matches against long, specific questions. Ask customers prompts that draw out the service, neighborhood and situation. Never write reviews for them; Google's policy prohibits AI-generated reviews and can suspend the profile (Agentive).
  4. Turn the Business Profile into a complete data feed. Every service, price range, service area, attribute and hour. Connect a booking link. Review Google's AI-drafted answers every week. This is the shortlist Google's agent calls from.
  5. Feed every listing pipeline, not just Google's. ChatGPT and Claude don't read your Google profile directly. Only 68% of business contact information on ChatGPT and Perplexity matches the Google profile (Search Counsel). Keep Bing Places, Apple, Yelp and the main directories identical.
  6. Be agent-ready. Publish prices or price ranges. Answer every call immediately, with an AI receptionist if needed, ready to quote and book. Make booking possible online and by text. The agent picks the business that can act.
  7. Buy the research moment, earn the decision. Use ChatGPT ads to reach people while they're comparing options, since money can't buy a place in the answer itself. Then make sure the organic answer names you when they're ready to hire.

All seven come down to one rule: be the business that machines can verify from many sources and act on immediately.

Where the billions are being wasted

The expensive mistakes share one flaw: they optimize the parts of the pipeline the machines barely use.

  • Mass informational content. Blog posts built to rank for "how to" questions are exactly what AI Overviews now answer on the results page, taking the click with them.
  • Link volume as the main deliverable. Links still feed rankings, but mentions predict AI visibility about three times better.
  • GEO gimmicks. Google says llms.txt files and special markup aren't needed, and the 10,000-business study above found weak or absent correlation for structured data, llms.txt and FAQ schema (Search Counsel).
  • Paying to get into the answer. OpenAI says ads can't influence ChatGPT's responses. Ad money buys the space beside the answer, not inside it.
  • Treating the website as the center. For local questions, the AI reads the business's own site for only a small share of its sources. The rest is what others publish.

How to measure it

Rankings stop meaning much when every answer is generated fresh. In one test, only 8 of 100 repeated questions drew on the same sources both times (PlusPoint). Measure like a pollster instead:

  • Share of answer. Run your 50 most important customer questions through Google AI Mode, ChatGPT and Claude many times a month. Track the percentage of answers that name you.
  • Google's own AI report. Search Console's generative AI performance report shows AI Overview and AI Mode appearances by page.
  • AI referral conversions. Give AI referrals their own channel in analytics and track leads, not sessions.
  • Agent calls won. Track how many calls from Google's agent and personal assistants turn into bookings.

The bottom line

The old game was being the best page. The new game is being the business that every source agrees on and every agent can act on. The companies spending billions on content and links are fighting over the part of the pipeline that's disappearing. The opening is in the sources AI reads, the sub-questions it asks, and the moment an agent picks up the phone.

One caveat: much of the research on AI search comes from SEO tool vendors and moves every few months. Treat these plays as hypotheses to test in your own market, and re-check them each quarter.

Sources