The recommendation layer: how AI picks which brands to name

Picture someone buying an espresso machine. Five years ago they typed "best espresso machine under $500," opened six tabs, and argued with themselves for a week before deciding. Now they ask ChatGPT and get three machines ranked with reasons, because the model has already read the reviews, compared the specs, and made the call for them. In Orbit Media's March 2026 survey of 1.1K US adults, 55 percent said AI chat was their primary or frequent tool for research.

Here's the honest part most posts skip: that same survey found Google usage hasn't declined, which means this is less a replacement story than an addition story. The shift is what people have always done when choice gets overwhelming: they borrow judgment from something they trust, and the trusted something just stopped being human. A new decision-maker showed up between your customer and your website, and it never scrolls, never gets bored, and has never once been charmed by a hero image.

That new reader is the problem I spend my working weeks on. I build AI visibility systems for enterprise brands, and this post is the mental model behind the work I publish at haydenschuster.com: what the recommendation layer is, how it behaves differently on every surface, and the loop I use to win it. And instead of hypotheticals, I'm going to show you the receipts.

01Who makes the decision now?

The searching didn't change much. The deciding changed completely.

Every row in that table used to be your customer's job. Search Engine Land calls this territory the AI decision layer, where AI weighs trust, relevance, and authority before deciding who makes the shortlist, and losing here means the customer never learns you existed.

The checkout row is the newest and, for ecommerce, the biggest. Google introduced the Universal Commerce Protocol in January 2026 with Shopify, Target, Walmart and other partners. In May, it announced Universal Cart checkout with Google Pay and Merchant Center AI performance insights, with the checkout features and reporting rolling out rather than available to every merchant at once. The direction is clear: when a shopper can buy from an AI answer, inclusion affects more than a click.

02What is the recommendation layer?

The recommendation layer is the space between a person's question and the answer an AI gives them. It's where ChatGPT, Gemini, Perplexity, Copilot, and Google's AI Mode fetch sources, weigh them, and choose which brands and pages get into the answer. In classic search your job ended at ranking, but here ranking is only the audition, and getting chosen is the job.

If you think your rankings have this covered, they don't. Ahrefs tested 15K prompts and found only 12 percent of links cited by ChatGPT, Gemini, and Copilot rank in Google's top 10 for the same prompt, and around 80 percent do not rank in Google's top 100 for the original query. Your rank tracker is watching one store while the model shops in several.

03One layer, five landscapes

"AI search" isn't one thing. Each surface retrieves differently, trusts differently, and rewards different work, and treating them as one blob is how teams end up optimizing for none of them.

The overlap figures in that table are Ahrefs data, and the spread is the point. AI Overviews cited Google's top-ranking pages much more often than the assistants in that study. But citation overlap does not reveal each provider's internal plumbing. Build for access, evidence and a good answer; then read the actual citations. One ranking report cannot speak for five surfaces.

04A mention is not a citation

A mention is the model saying your name, while a citation is the model pointing at your content as the source. Here's the shape of each, using a real page from CheckMyTap, one of my sites.

Most brands celebrate the first one and stop. Screenshots are not a strategy, and the two don't reliably travel together:

A real recommendation is both at once: the model says your name and uses your content as the proof. Being named without being cited means the model is repeating your reputation rather than your facts, and being cited without being named means you did the work and got none of the credit.

05The three tests your content has to pass

Every AI answer runs your content through three checks, in order, and failing one means the rest don't matter. If you know my Found, Understood, Chosen framework, this is the same system pointed at a new judge: visibility is Get Found, verifiability is Get Understood, quotability is Get Chosen.

1. Visibility: can I even get to this?

Visibility means the model can reach and retrieve your content when it goes looking, which makes it a question about retrieval rather than rankings. If robots.txt blocks AI crawlers they can't read you, and if your words only appear after scripts load, some systems see an empty page, which is why my technical foundations work starts with what survives a raw crawl. The model also doesn't limit itself to your site; it leans on the places it already trusts, like review platforms, forums, Reddit, and YouTube, so sometimes the fastest win isn't on your domain at all.

Pass it when AI crawlers are allowed on purpose, your key content lives in the raw page code, and you show up on the third-party sources models pull from in your category.

2. Verifiability: can I trust this enough to repeat it?

Verifiability means the model can check your claims before repeating them. Think of it as the world's most nervous librarian: repeating something wrong is expensive, so it favors content it can trace. This is where your moat does the work, because the data only you have, assembled the way only you assembled it, is the one claim a model can verify against you and nobody else. It's why CheckMyTap opens every city page with that city's actual readings, and why modern search systems need consistent, checkable facts about who you are everywhere they look. The original GEO research from Princeton, Georgia Tech, and IIT Delhi found that citing credible sources inside your own content was a top tactic for AI visibility, which is worth reading twice: citing others makes you more citable.

Pass it when claims link to named sources, content carries a real byline instead of "admin," and your site, profiles, and listings all agree on who you are.

3. Quotability: can I lift a clean answer out of this?

Quotability means one of your sentences still makes sense when lifted into an answer. The Princeton-led GEO study built a 10K-query benchmark and evaluated tactics on its 1K-query test split. In its experimental answer engine, statistics, quotations and cited sources improved visibility, with the best result near 40 percent on one measure; keyword stuffing did little. Your customers have written quotable material too. A review with a name and a specific experience can say more than a line of marketing copy. The difference is easier to see than explain:

Pass it when every section opens with a complete, specific, standalone claim, which is the whole idea behind answer-first structure. The first sentence above contains zero facts a model could repeat; the second gives the reader a sourced finding they can check.

06What the research says works

The GEO paper (Aggarwal et al., KDD 2024) evaluated content tactics on a 1K-query test split of its 10K-query benchmark. In that experimental setting, adding verifiable detail and quotations improved its visibility measures.

Notice what this post is doing. The answer sits at the top, every stat names its source, and the highlighted sentences are the ones a machine could lift whole. That's the three tests applied to the post about the three tests.

07GEO is SEO

Time to kill the laziest myth in the industry while we're here.

Look back at everything above and notice what it's made of. Crawlability and rendering are technical SEO. Entities, structured data, and consistent facts are the understanding layer SEO has been building for a decade. Answer-first content matched to real intent is what good SEO content always was, and it's the whole reason Found, Understood, Chosen maps onto this so cleanly. GEO is SEO. The fundamentals didn't change; the judge did. So no, SEO is not dying. It has never been more essential, because the machine now reads everything before your customer reads anything, and the discipline that decides what machines find, understand, and choose is the one this industry has been practicing all along.

What did change is that it stopped being a solo sport. The models lean on Reddit threads, reviews, and YouTube, which is social and community's turf. Retrieval depends on rendering, structured data, and index pings, which is the developers' turf. The moat data comes from strategy and product, and the measurement loop touches analytics. The brands winning this run SEO as a cross-functional system rather than a department, because if your SEO team, your developers, and your social team don't share a roadmap, the model meets three different versions of your brand and trusts none of them. The machines will answer either way. The question is whether they answer from verified sources or from noise, and making sure it's the former is the job. That's why this work matters more now, not less.

08The loop I run for enterprise brands

Every enterprise program I run settles into the same four-beat loop, and it works at any size.

Listen. Run the real questions buyers ask, on a schedule, across ChatGPT, Gemini, Perplexity, and Google's AI surfaces, and track two things separately: whether the brand is in the answer, and whether the brand's own content is the source. A brand the model describes is at the mercy of its loudest reviewer, while a brand the model cites is telling its own story.

Mine. The dashboards are a content calendar wearing a scorecard costume. Every question, sub-search, and grounding query the brand hasn't directly answered becomes a brief.

Answer. Rewrite the pages that matter so each section opens with the answer, turn the mined sub-questions into FAQ entries that actually resolve them, and publish the moat data only this brand has.

Verify. Watch mentions and citations move separately on each surface over the following weeks. When citations rise but mentions stay flat, the content is trusted and the brand framing needs work; when mentions rise without citations, the brand has fame and no receipts. The split tells you what to fix next, winners get the next round of investment, and the loop starts again.

Every green box in this post is that loop running on my own hobby site, with Bing's grounding report as the citation scoreboard. The Quack Report runs the same playbook on Oregon football history, keeping every rivalry record and season result at its own citable address, and CheckMyTap runs it on city water data. The subjects have nothing in common, but the three tests are identical.

09Your dashboards are a content strategy

Most teams read these four tools as report cards, but the better read is that each one hands you the next thing to write. It helps to know why measurement is hard in the first place: someone asks ChatGPT, gets your brand, then googles your name, and your analytics logs a branded search. The AI's fingerprints are nowhere in the data, so you triangulate.

Peec AI: what to answer next

Feed Peec AI your buyer questions and it runs them across ChatGPT, Perplexity, Gemini, and Google's AI surfaces on a schedule, tracking whether you appear, how you're framed, and which sources shaped each answer, and it exposes the fan-out queries behind each one.

The move: Those fan-outs and tracked questions are a ready-made FAQ outline. Answer each one directly on your site and you're covering demand your keyword tools will never show you.

Blind spot: Scheduled test prompts, not real user sessions.

Bing Webmaster Tools: what retrieval already likes

Microsoft's AI Performance report, introduced in February 2026, shows page-level citations across Copilot, Bing AI answers and select partners, plus a sample of grounding queries. It's a free look at where pages were cited and what the system was trying to answer, and it's the source of the citation numbers in this post.

The move: Treat grounding queries as briefs. Read the cited page and the question together, then improve the answer or build the missing page. Test ChatGPT separately; Bing citations do not tell you which sources ChatGPT used.

Blind spot: Microsoft's ecosystem only. No click data yet.

Google Search Console: how people phrase it now

Google folds AI Overviews and AI Mode into the standard performance report, so you read it through patterns. Impressions climbing while clicks stay flat deserve an investigation, not an instant AI win label: inspect the queries and live results. On the fan site, one character page pulled 87.9K impressions in three months at a 0.35 percent click rate. That proves demand and thin click-through; the result page tells you where the answer went. My approach to using GSC as a visibility dashboard matters more now, not less.

The move: Mine the query report for the longer, question-shaped queries showing up against each template. Those are prompts in disguise, and they track exactly how search behavior is shifting. Give every one a direct on-page answer.

Blind spot: No dedicated AI report. You're inferring from blended data.

GA4: which pages convert the machine's clicks

GA4 added an AI Assistant channel in May 2026. It groups recognized assistant referrers; Google's channel definition excludes AI Overviews and AI Mode, so I still inspect session sources and use a custom group when I want to include more referrers. On the fan site, the AI channel catches about a hundred sessions a month (155 in the 30 days to July 17), which is real but a fraction of what the citation data shows upstream. Seer Interactive's single-client case study measured ChatGPT referrals converting at 15.9 percent against 1.76 percent for Google organic from October 2024 through April 2025. That comparison describes one client and one period; it does not predict the conversion rate for another site.

The move: Run landing page by AI source. The pages already earning AI clicks and conversions are your template, so make the rest of the site look like them.

Blind spot: Undercounts hard. Many AI visits arrive with no referrer and get logged as Direct.

10Bringing it together: the build sheet

Three tests, the real assets that pass them, and the data that finds your priority one work. This is the sheet the whole post has been building toward.

That's the priority rule, and it holds across every program I run, whatever the industry: P1 is wherever a surface already wants you but can't use you. That means grounding events that never become citations, mentions that never come with citations, and questions sitting in your own data that nothing on your site answers. Fix the pull before you build more push. This is the same system behind my modern search portfolio work, just written down.

The companion rule matters just as much: a winner is a map for the next build. When a page starts earning citations, the data clustering around it (the fan-out phrasings, the grounding queries, the question-shaped searches) is the machine handing you the outline for the next FAQ entry, the next post, and the next topic cluster. Prioritize those neighbors and double down, because the surface has already voted, and expanding a proven winner beats guessing at a new one every time. It's why the fan site's release date winner sits inside a whole books cluster: reading order, recaps, a page per book, each catching another question from the same pool of demand. Gaps tell you where to start. Winners tell you where to build.

11Where do you start?

In this order, because each step is cheap and each one makes the next one smarter.

1. Run your own answers, manually, today. Put your 20 real buyer questions through ChatGPT, Gemini, and Perplexity and screenshot everything, which gives you a free baseline showing whether you're mentioned, cited, both, or invisible.
2. Check that the machines can reach you. Review robots.txt for AI crawler rules you didn't know you had, and load key pages with scripts off. If the content disappears, so do you.
3. Rewrite your most important pages answer first. Open every section with one standalone, specific sentence, and put a number in it wherever the truth allows one.
4. Make yourself checkable. Link claims to sources, put real names on content, and make every description of your brand agree with itself.
5. Then build the stack and run the loop. Use Peec for answers, Bing for citations, GSC for phrasing, and GA4 for the money end, and run the loop: listen, mine, answer, verify.

12What decides who AI recommends

The decision moved. People still search, but AI answers now help them read and compare before they reach a site. The surfaces behave differently: Google's AI Overviews overlap heavily with top results in Ahrefs' study, while Google's AI Mode can fan out into related queries. The assistants need their own citation checks. With agentic checkout rolling out through UCP and Universal Cart, the answer is starting to double as the store shelf.

Being named is awareness and being cited is trust; a real recommendation is both at once, and the data says the two rarely travel together on their own.

Three tests decide everything: whether the model can reach you, verify you, and quote you. GEO is SEO pointed at a new judge, your dashboards hand you the roadmap when you read them as inputs instead of report cards, P1 is wherever a surface already wants you but can't use you, and every win maps the next one. Bing's October export counted 6.7K citations against one Book 9 release-date question on my hobby fan site. The excuse shelf is empty.

The ten blue links gave everyone a shot, but the recommendation layer only has room for the sources it trusts. Be one of them.