Copy for Originals · Guide · AI search and recommendations
In October 2026 we published seven studies and guides on how AI assistants read websites, choose between options and recommend companies, and read more than 200 sources behind them. This page distills that work into eight findings, a four-gate model of how an AI chooses, and seven steps a company can take in order. The findings are in the first half. The steps are in the second. Nothing here is a guarantee, because the evidence says nobody can give one.
AI assistants are a research layer with a person still deciding, and whether they name a company comes down to four gates. Can the assistant read your pages? Does it already know your name from the rest of the web? Can it verify a specific claim about you, such as a price, a date or a number? And does it prefer you when the options are close, which depends on ratings, third-party sources it trusts, and position, not on your copy. Rankings inside a single answer are noise; how often you appear across many answers is stable. Most of what is sold as AI optimization, including schema markup, llms.txt files and FAQ blocks, shows no measurable effect on citations. The steps with evidence behind them are cheap, and most are things a company already controls.
Eight things held up across our own studies, the peer-reviewed and preprint experiments, and the platforms’ own documents. Each one below starts with the finding, then the strongest numbers behind it. Where a number comes from a company that sells AI visibility tools, we say so.
Finding 1
Helping them research. Every survey that separates the two puts AI-assisted research around five times ahead of delegated buying, and the one product that tested autonomous buying at scale was wound down. IBM and the National Retail Federation found 45% of about 18,000 consumers in 23 countries use AI during the buying journey, and every breakdown was research: comparing products, reading reviews, hunting deals. Gartner found 11% would let AI decide even a low-stakes purchase. DOSS found 6% would let an agent buy on its own. Checkout.com’s merchants reported about 3% of transactions involving an agent.
OpenAI launched Instant Checkout inside ChatGPT in September 2025 and moved checkout back to merchants’ own sites by March 2026. Walmart said conversion inside the chat ran at about a third of the rate of clicking out to its own site. The platforms have since converged on one shape: discovery in the AI, checkout on the merchant’s site, and ads beside the answer. ChatGPT’s ads reached a reported $1 billion annualised run rate in under 200 days.
What it means. The person who reads an AI answer is the same person who later searches your name, visits your site and decides. AI answers shape the shortlist. They rarely close the sale. Full detail is in our review of the evidence on AI shopping agents.
Finding 2
Because the answer is sampled, not looked up, and the sample changes with every run, every rewording and every model version. In SparkToro’s test of 2,961 runs, two answers to the same question listed the same brands fewer than once in 100 runs, and in the same order about once in 1,000. The University of St. Gallen found identical prompts run within 24 hours shared only 32 to 43 percent of their cited sources, and that at least seven runs per prompt are needed before a brand’s share settles. Rewording matters more than rerunning: one reanalysis of 190,000 observations found reliability of 0.96 across 250 questions run five times each, but 0.27 across five rewordings of one question. A model upgrade can move a product from 45% to 77% of an agent’s choices with no change on the website, as the Columbia and Yale ACES experiments found when Claude Sonnet 4 was replaced by Opus 4.5.
The average is stable even when the single answer is not. The same SparkToro data showed four headphone brands appearing in 55 to 77 percent of 994 answers although the exact list almost never repeated. SparkToro’s authors concluded that a visibility share across dozens or hundreds of prompts, each run several times, is a reasonable metric.
What it means. “We’re number two in ChatGPT” is a coin flip. How often you are named across many prompts and runs is a measurement. Any report that gives you a rank in an AI answer is reporting noise.
Finding 3
Because ranking and being cited are different steps. In March 2026, only 38% of the pages cited in Google AI Overviews ranked in the top ten for the original query, down from about 76% in July 2025, according to Ahrefs. An assistant answers one question by running several shorter background searches of its own (“query fan-out”), and it reads what those searches return. OpenAI’s own documentation gives the example of “good restaurants near me” becoming “top restaurants San Francisco”. Across all categories, 95 to 99 percent of those background searches have no search volume in any keyword tool, and only about 27% recur from run to run, in vendor measurements.
Cited pages do match the question more closely. In 1.4 million ChatGPT prompts, Ahrefs found cited pages’ titles were closer to the prompt than non-cited pages’ titles (cosine similarity 0.602 against 0.484). In about 2.1 million citations across four AI systems, how closely the page content matched the prompt was the strongest page-level factor. Pages ranking first in Google were cited 3.5 times more often than pages outside the top 20, in AirOps data.
What it means. Ranking still helps, because it puts you in the pool. But the page that gets cited is the one that answers the sub-question the assistant actually searched for, in words close to the question. Google’s own guide says optimizing for its AI features is still SEO, and for Google’s surfaces that is broadly true. It says nothing about how ChatGPT, Claude or Perplexity fetch pages.
Finding 4
Often not, and the owner usually doesn’t know. In our study of 100 Chicago personal injury law firm websites, 13 of the 78 sites that could be tested refused a request carrying the name GPTBot, ClaudeBot or PerplexityBot in all five runs. 12 of the 13 had no robots.txt rule against that crawler, and 11 of the 13 served the same request every time when it carried Googlebot’s or Bingbot’s name. The refusal came from a firewall, a hosting provider or a security plugin, not from a decision anyone recorded. A separate audit of 200 UK and US retailers found 28 refused an ordinary browser agent, every one through bot defenses rather than stated policy; Walmart, which wants agents, admitted blocking one by mistake.
A page the crawler can fetch can still be unreadable. None of the crawlers from OpenAI, Anthropic, Meta, ByteDance or Perplexity run JavaScript, according to Vercel’s analysis of crawler traffic. Only Googlebot and AppleBot render it. Text inserted by a script, prices computed in the browser and numbers drawn in images cannot be cited. Main-content extractors also discard hidden nodes, so an answer inside a collapsed accordion may be dropped.
What it means. Before any content work, confirm that the services can read you. The check takes an afternoon and is described in step 1 below. The full method is in do law firm websites block AI crawlers? and the pipeline is in how to write content that LLMs retrieve and cite.
Finding 5
Relevance to the question and specific, verifiable facts. Not layout. The controlled experiments agree, and they disagree with most of the advice being sold.
| Lever | What the evidence found | Verdict |
|---|---|---|
| Explicit prices, numbers and recent dates | In 252,000 controlled trials across six language models, explicit prices and recent timestamps consistently increased selection; organised sections against dense paragraphs “had no impact”. LLM rerankers also promote passages that look recent. Strong | Do it |
| Close match to the question | Strongest page-level factor in 2.1 million citations; cited titles closer to the prompt in 1.4 million ChatGPT prompts; prefixing a page with the question it answers substantially raised how often models relied on it (ACL 2024). Moderate | Do it |
| Statistics, quotations and cited sources | Raised a source’s visibility by up to about 40% in the original GEO study (KDD 2024), but that was a 2023-era model on informational questions. A NeurIPS 2025 benchmark of the same tactics found significant improvement in only 3 of 54 cases, and adding statistics lowered rank in 19 of 24 settings. Mixed | Do it for the reader, not for the engine |
| Schema markup (JSON-LD) | 1,885 pages that added schema, against matched controls: −4.6% in AI Overviews, +2.4% in AI Mode, +2.2% in ChatGPT, the last two indistinguishable from zero (Ahrefs). None of five AI systems found a price that existed only in JSON-LD. Moderate | Keep for classic search; never the only place a fact lives |
| llms.txt | 97% of the files on 137,210 domains received no requests at all in May 2026 (Ahrefs); no relationship with citations across about 300,000 domains (SE Ranking); Google Search ignores it. Moderate | Skip |
| FAQ blocks and answer-first formatting | Small positive association that was not significant after controls; Google retired FAQ rich results on May 7, 2026 and says there is no requirement to break content into tiny pieces. Bing says Q&A pairs can be lifted word for word, with no data. Weak | Useful because FAQ questions match real queries, not as a lever on its own |
| Keyword stuffing, superlatives, emotional copy | Keyword stuffing did worse than nothing (−8.7%) in the GEO study. In a 2026 test across five engines on 13,747 product queries, “authoritative tone” was negative on all five, and superlatives and emotional copy were flagged 99 to 100 percent of the time under a simple defence. Strong | Stop |
What it means. Write the page for a careful reader who wants a specific answer with a number in it, and make sure the number is in the HTML. That is the whole of the content advice the evidence supports. Our evidence-rated guide rates every common practice.
Finding 6
Specs when a spec settles it, and salience when it doesn’t: ratings, the platform’s own badges, position in the list, and sources the model already trusts. In the lab, agents are more persuadable than people, by different cues.
What it means. You cannot write your way past a lower rating or a higher price. You can make sure the agent can verify the attribute you win on, and you can be in the sources it trusts. All of this is from sandboxes with 2025 and 2026 model versions; the directions replicate across three independent labs, the magnitudes change with every release.
Finding 7
For a question that names your brand, mostly yours. For a question that doesn’t, mostly other people’s. In one B2B dataset of 57 million citations, brand-owned pages were 30.7% of citations on prompts that named a brand and 2.24% on prompts that didn’t; 85% of unbranded answers cited no brand-owned source at all. In 17,524 “best product” answers across five engines, 70% cited an affiliate review site. In healthcare, directories such as WebMD and Vitals carry 52.6% of citations across ChatGPT, Gemini and Perplexity, and in a peer-reviewed test of 40,500 requests for an orthopedic surgeon, ChatGPT declined to name anyone 52.8% of the time. All but the last of those figures come from vendors.
Being present in the retrieval path is the difference between appearing and not. In one observational study of 34,960 answers, a brand with no page in the retrieval path was named 2.8% of the time by GPT and 3.8% by Gemini; with its own domain cited, 49.0% and 58.4%. Across 75,000 brands, how often a brand is mentioned on the web correlated with its AI Overviews visibility at 0.66; for ChatGPT the correlation with web authority metrics was very weak. On Reddit, practitioners named third-party mentions as the way to get cited more often than anything else (444 replies), ahead of content on your own site (343), in our review of 1,073 AEO threads.
What it means. Your own pages decide how you are described once you are named. Other people’s pages decide whether you are named. No controlled study has shown that buying mentions raises your recommendation rate, and Google says inauthentic mentions are not as helpful as they look.
Finding 8
Exposure on Google can be observed. Exposure in the chat assistants can be sampled with a stated error. The influence that never clicks can only be estimated. Google’s Search Console generative AI report, available to every site since August 31, 2026, shows impressions in AI Overviews and AI Mode by page, country, device and date. It shows no clicks, no queries, no split between the two features and has no API. Bing Webmaster Tools has shown Copilot citations and a sample of the search phrases the AI actually ran since February 2026. ChatGPT, Claude, Gemini and Perplexity publish nothing; 45 of the 308 Reddit threads we coded said so outright.
The sampling tools fill that gap with their own prompts. Prices run from $29 to $489 a month, and the Reddit threads about them (55 of 308) object that the prompts are synthetic, that API answers differ from the app, and that two tools disagree. The maths is unforgiving: telling a 20% mention rate from 30% takes about 290 answers per group, and a single prompt asked three times with one mention has a 95% interval of roughly 6% to 79%. A three-prompt audit is not a measurement. The direct traffic is small: AI referrals were 0.28% of visits across 74,752 sites in Ahrefs’ data and 0.32% across 101,574 in SE Ranking’s, three-quarters of it from ChatGPT, and mobile apps and pasted links lose the referrer. The larger effects leave no trace: in the only randomized trial, AI Overviews cut organic clicks by 38% on the queries where they appear, and a vendor panel found an AI recommendation raised same-name Google searches by 4.3 percentage points among people with no prior engagement.
One deadline may change this. The UK Competition and Markets Authority has required Google to give publishers AI impressions, clicks and click-through rate, disaggregated from the rest of search, by about December 3, 2026. Whether that data reaches sites outside the UK is not stated; the impressions report did go worldwide after a UK start.
What it means. Report AI exposure the way broadcast reach is reported, not the way paid search is. Our Reddit study ranks the twelve measurement problems practitioners hit, and step 6 below says what to track.
The eight findings reduce to four gates, in order. A company that fails an early gate cannot fix it at a later one, and most of what is sold as AI optimization works on gate 3 alone.
| Gate | The question the system is effectively asking | What decides it | Who controls it |
|---|---|---|---|
| 1. Readable | Can I fetch and parse this company’s pages? | Firewall and bot rules, robots.txt, JavaScript, hidden content, presence in Google’s and Bing’s indexes | You, entirely |
| 2. Known | Does this name come up when I search the sub-questions, and do I already know it? | Mentions across the web, listings and directories, reviews, trade and editorial coverage, search interest, ranking for the background searches | Mostly others; you can earn it |
| 3. Verifiable | Can I quote a specific fact about this company that answers the question? | Prices, dates, numbers, who it is for, in plain HTML, in sections that stand alone and match the question’s wording | You, entirely |
| 4. Preferred | Among the options that pass, which do I pick? | Ratings and review counts, winning on a measurable spec, sources the model trusts, position and platform badges, the user’s stated criteria; then randomness | Partly you, partly the platform, partly chance |
Two things follow. First, the gates you control fully (1 and 3) are the cheapest, and most companies have not checked them. Second, gate 4 has a floor of randomness that no amount of work removes, which is why appearance share across many runs is the only honest score.
Each step names the gate it opens, what to do this week, and the evidence it rests on. The first three cost time rather than money. Do them in order; a later step done before an earlier one is wasted.
Open yoursite.com/robots.txt and search for OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User and PerplexityBot. A User-agent line naming one followed by Disallow: / asks it to stay out. Then ask whoever runs your hosting, your Cloudflare account or your security plugin whether a bot rule refuses AI crawlers, because 12 of the 13 refusals in our study did not appear in robots.txt. Ask for server log lines containing those names and the status code each got: 200 means the crawler received the page, 403 or 429 means it didn’t. Fetch your most important page with curl and read the raw HTML; anything missing from that output cannot be cited. Decide training and search separately: blocking GPTBot and ClaudeBot keeps you out of model training and does not remove you from AI answers; blocking OAI-SearchBot removes you from ChatGPT’s. Submit your sitemap to Bing Webmaster Tools as well as Google Search Console, because Copilot answers from Bing’s index. Our free tool, Can AI read my site?, runs most of these checks for you in about a minute.
For each thing you sell, one page that states the price or price range, what is included, who it is for and who it isn’t, how long it takes, and what can go wrong, in text, not in a PDF, an image or a script. Open each section with the answer and name the subject instead of “it”, because sections are retrieved one at a time. Use real tables with header cells. Show a visible “Last updated” date and change it only for real changes. Write headings in the words a customer would type, including the long ones, because those are the assistant’s background searches. This is the step the controlled experiments support most strongly: explicit prices and recent dates increased selection in 252,000 trials, and layout alone did nothing.
Take one question your customers ask before they buy that no one has answered with numbers, and answer it with data you collected: what things actually cost in your market, how many competitors do X, what customers complain about most, measured and dated. This is the content that gets quoted because it is quotable, that gets mentioned because it is new, and that a competitor cannot produce in a minute with the same tools you have. Every study on this site was made this way; the method is described on our SEO and AEO service page. Text that says what everyone else says has no reason to be placed ahead of the others and nothing in it to quote.
List the five to ten places an assistant reads when someone asks for a provider or product in your category: the directories, the comparison sites, the trade publications, the “Wirecutter” of your field, and the forums where real customers talk. Ask the assistants yourself and note which sources they cite. Make sure your listings on those sites are complete, consistent and current, with the same name, address, categories and prices as your own site. Earn reviews, because ratings are the strongest visible lever and a 0.1-star edge can erase a bigger brand’s advantage. Pitch the publications with the original data from step 3, which is the only thing a journalist can’t get elsewhere. Do not buy mentions, plant threads or fabricate badges: no controlled study shows bought mentions work, Google says inauthentic ones aren’t as helpful as they look, superlatives are flagged 99 to 100 percent of the time, and fabricated claims are fraud.
Agents punish any option that loses on a measurable spec and reward a verifiable edge far more than a claimed one. Pick the attribute you genuinely win on in your market, price, speed, opening hours, a rating, a guarantee, a certification, and state it as a plain fact with a number and a date on the page from step 2 and in your listings from step 4. “Open Saturdays until 4 pm” beats “convenient hours”. “4.8 stars from 312 reviews on Google” beats “highly rated”. Leave out the adjectives; the models increasingly filter them, and they can backfire.
Track four numbers monthly. Google AI exposure: the Search Console generative AI report, impressions by page. Bing: the AI Performance report, citations and the grounding queries it shows. Direct AI referrals in your analytics (chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, and the utm_source=chatgpt.com tag), as a floor, plus “how did you find us?” on every form and call. Branded search demand over time, because that is where AI influence shows up. If you sample the assistants yourself or buy a tool, insist on unbranded prompts seeded from real queries, at least seven runs per prompt, two-to-four-week windows, share of answers with a confidence interval, and no rank and no sentiment score. Store the raw answers and dates yourself so a tool change doesn’t erase the series. State once, in writing, that the unclicked share of AI influence and the assistants’ prompt volumes cannot be measured with current tools. If you have UK users, have your Search Console exports and branded-query filters ready for December 3, 2026.
AI answers favour the already familiar. Four headphone brands appeared in 55 to 77 percent of answers; known brands win every tie; buyers choose from their day-one shortlist about 90% of the time in B2B studies. The best-supported budget split in consumer markets is about 60:40 brand to activation, from the IPA databank, and the evidence that preference is built over years points the same way. Specs belong in activation and product data, where the agents read them. Report AI the way you report reach: share of answers, branded-demand lift, direct referrals as a floor. Expect paid placement beside AI answers to grow, because it is the only AI surface with full measurement, and measure its incrementality before following the budget there.
Each item below is sold as AI optimization. Each one has either no measured effect, a measured negative effect, or a measurement problem that makes the deliverable meaningless.
Not as far as anyone has shown. ChatGPT has sold ads since February 2026 and Google sells ads inside AI Mode, but both say the ads do not change the answer, and no platform has disclosed any mechanism where paying changes an organic recommendation. In the Columbia and Yale ACES experiments, a “Sponsored” label lowered an AI agent’s selection of a product from a 10% baseline to about 8%. Ads inside AI answers are also closed to health, finance and politics on both Google and ChatGPT.
Not measurably. Ahrefs tracked 1,885 pages that added schema markup against matched controls and found citations changed by −4.6% in Google AI Overviews, +2.4% in AI Mode and +2.2% in ChatGPT, the last two indistinguishable from zero. Ahrefs also found that 97% of llms.txt files on 137,210 domains received no requests at all in May 2026, and SE Ranking found no relationship between the file and citations across about 300,000 domains. Google retired FAQ rich results on May 7, 2026, and its guide says there is no requirement to break content into tiny pieces. What does have evidence is relevance to the question, specific facts such as prices, numbers and dates, and being readable in raw HTML.
Because the answer is sampled from a distribution, not looked up. In SparkToro’s test of 2,961 runs, two answers to the same question listed the same brands fewer than once in 100 runs, and the same brands in the same order about once in 1,000. The University of St. Gallen found identical prompts run within 24 hours shared only 32 to 43 percent of their cited sources. The average is stable even when the single answer is not: how often a brand appears across many prompts and many runs is a reasonable metric, while its position in any one answer is noise.
Because ranking and being cited are different steps. In March 2026 only 38% of the pages cited in Google AI Overviews ranked in the top ten for the original query, according to Ahrefs. An assistant answers one question by running several background searches of its own, 95 to 99 percent of which have no search volume, and it reads the pages those searches return. ChatGPT, Claude and Perplexity also fetch pages with their own crawlers, which a firewall can refuse and which do not run JavaScript, so a page can rank and still be unreadable to them.
Check whether the AI services can read the website. In the Copy for Originals study of 100 Chicago law firm websites, 13 of 78 testable sites refused a request carrying the name GPTBot, ClaudeBot or PerplexityBot in every run, and 12 of those 13 had no robots.txt rule that showed it; the refusal came from a firewall or security plugin. Open yoursite.com/robots.txt and search for the crawler names, ask whoever runs the hosting or Cloudflare account whether a bot rule refuses AI crawlers, and ask for server log lines containing those names with the status code each one got. A 200 means the crawler received the page.
This page synthesizes seven pieces of work published by Copy for Originals in October 2026. Each has its data, sources and limits on its own page; the sources this page relies on directly are listed below.
Read at source between October 13 and 16, 2026. Sources from companies that sell AI visibility, listings or tracking products are marked (vendor).
How to cite: Copy for Originals, “What decides whether an AI recommends you, and what to do about it,” October 17, 2026, https://tools.copyfororiginals.com/what-decides-ai-recommendations/.