How AI Decides Which Brands It Recommends

Shoppers increasingly ask an AI assistant “what should I buy?” and act on the two or three brands it names. Here is how that shortlist is actually assembled — and what a brand has to fix to be in it.

Some brands we cover are clients of our publisher. What we recommend is decided by our editors, and prices are read on the date noted in each article. Our standards.

Search did not die. It moved into the conversation.

For twenty years, the shape of online discovery was fixed: type a few keywords, get a page of links, click two or three, decide for yourself. Every part of that flow assumed a human doing the comparing.

That assumption is quietly dissolving. A shopper now types a full sentence — “what’s the best cast iron skillet for an apartment stove under $80?” — and gets one answer back, in prose, with reasoning attached and one to three products named. The comparison work that used to happen across a dozen open tabs happens before the shopper ever lands on a website.

This is not a prediction. Research firm Gartner projected that traditional search engine volume would fall 25% by 2026, with that demand absorbed by chatbots and virtual agents. The measured behavior has followed. According to Adobe Analytics data on AI-referred traffic, visits arriving at U.S. retail sites from AI sources were up 393% year over year in the first quarter of 2026 — retail’s growth outpacing travel, financial services, media and software. In May 2026, Adobe found those AI-referred visitors converted at a rate 54% higher than visitors from non-AI sources, and generated more revenue per visit. A year earlier, the same cohort converted at roughly half the rate of everyone else. That reversal is the whole story in one number: AI referrals stopped being idle browsing and started being purchase intent.

Editors’ note: we cover this because it changes what a recommendation is worth. When an assistant names three brands instead of listing thirty, the difference between being named and being invisible is no longer a ranking position — it is binary.

Four steps, from links to agents

It helps to see the shift as four stages rather than one event, because most brands are currently optimized for the first and being evaluated by the third.

1. Traditional search. Keywords in, ten links out. Brands competed for position, and the shopper did the synthesis.
2. AI search. Full questions in, one synthesized answer out. The system reads sources, filters them, and explains differences in its own words.
3. AI shopping assistants. The conversation gets transactional. Assistants compare specific products, hold context across follow-up questions, and recommend named brands with stated reasons.
4. AI purchasing agents. The agent completes the checkout. Agentic storefronts are already transacting, which means product data has to be correct enough for software to buy from it unattended.

Each stage moves one more human judgment inside the machine. In stage one, a shopper decided which link to trust. By stage four, that trust decision is made from structured data, third-party evidence, and whatever your product page can prove.

How a recommendation actually gets made

Strip away the mystique and the sequence is fairly mechanical.

A question is asked. Not “grip socks” but “which grip socks stay put during barre if my feet sweat?” The query carries constraints, and constraints are what get matched against product attributes.

Sources are evaluated. The system pulls what it can read: product pages, structured data, reviews, editorial coverage, forum threads, spec sheets. It weighs them against each other and drops anything it cannot confirm. This is the layer where visibility is won or lost, and it is largely invisible from a marketing dashboard.

A shortlist gets named. One to three brands, usually with a reason attached — “this one is heavier but holds heat longer.” The reason is quoted from something the machine read. If nothing on your site or in third-party coverage gives it a reason to name you, it names someone else.

The site visit happens late. The shopper arrives already convinced, deep in consideration, often ready to buy. That is precisely why AI-referred conversion rates look the way they do.

What Google itself says about optimizing for this

There is a lot of noise in this space, and it is worth anchoring on the primary source. Google’s own documentation on AI features and your website is unusually blunt: the best practices for SEO remain relevant for AI Overviews and AI Mode, there are no additional requirements to appear in them, and no special optimizations are necessary. AI features are rooted in the same underlying search systems, not a separate index you can game.

Two things follow from that, and they pull in opposite directions.

The first is that nobody can sell you a backdoor into an AI answer. Any pitch that implies one is selling something that does not exist.

The second is that “the fundamentals still apply” is not the same as “you are already doing enough.” The fundamentals now include work most sites skipped for years because a human reader could paper over the gaps: pages that render without JavaScript, product attributes written out instead of shown in a photo, prices and availability marked up so they can be read without guessing, consistent naming so the brand resolves to one entity rather than three near-duplicates. A person can infer that “Sizes: see chart” means something. A model reading your page cannot.

Eight things worth auditing this quarter

If you run a brand, this is the shape of the actual work. None of it is exotic, and most of it is measurable.

1. Visibility baseline. Run your real buying prompts — the sentences customers would actually type — across the major assistants and record who gets named. You cannot improve a number you have never measured, and the answer is often “a competitor, consistently.”
2. Machine readability. Check what a crawler receives, not what a browser paints. Blocked bots, client-side-only content, and interstitials that hide specs are all silent disqualifiers.
3. Entity clarity. One canonical brand name, one category, consistent founder and origin facts everywhere they appear. Ambiguity is how a brand gets confused with something else, or dropped for safety.
4. Structured data. Implement and validate Product, Offer, review and FAQPage markup so price, availability, ratings and specs are stated as data, not implied by layout.
5. Attribute-level product context. Materials, dimensions, compatibility, care, sizing, use cases, and what the product is not for. Comparison requires attributes; missing attributes mean exclusion from comparisons.
6. Content in the shape of the question. Buying guides, honest comparisons, and structured FAQs written the way shoppers ask, rather than the way brand teams talk.
7. Third-party evidence. Most brand mentions in AI answers trace back to pages you do not own — reviews, roundups, community threads. Presence there is not a vanity metric; it is the evidence layer.
8. Continuous monitoring. Models, prompts and shopping surfaces change monthly. A one-off audit ages badly; tracking mentions, citations and share of voice over time is what turns this into a program.

Why the checklist keeps ending up in the same place

Read that list again and notice what it has in common: every item is about removing ambiguity. Not persuasion, not creative, not spend. An AI answer is a compression of everything a system can verify about a category, and compression discards whatever is unclear. Brands that lose in AI answers are rarely worse products. They are usually less legible ones.

That is also why the work is unglamorous and cumulative. There is no campaign that fixes it. It is a sequence — audit what happens today, fix what machines cannot read, make the brand resolve cleanly as an entity, state the facts as data, write the pages that answer real questions, enrich the product detail, earn third-party evidence, then keep watching. Teams that treat AI search optimization as an ongoing operating discipline tend to compound; teams that treat it as a one-time project watch their visibility drift as models update.

What this looks like on a small budget

The list above can read like an enterprise program, and it can be one. It does not have to start that way. If you sell fifty products and have no dedicated marketing team, the order of operations matters more than the volume of work.

Start with the baseline, because it is free and it settles arguments. Write down the ten sentences a customer would actually type before buying from you, ask them, and save the answers. That single document tells you whether you have a visibility problem, an evidence problem, or a category problem — and it tells you which competitors keep getting named in your place, which is uncomfortable and useful.

Then fix your best-selling product page completely rather than every page partially. Full attributes, correct markup, a real FAQ built from the questions your inbox already receives, and a plain description of who the product is wrong for. Once one page is genuinely legible, cloning that structure across the catalog is a template exercise rather than a strategy exercise.

The piece small brands underrate is the evidence layer. An assistant that cannot find anyone besides you talking about your product has very little to weigh, and it tends to reach for brands with a paper trail. Getting reviewed honestly — including by outlets that will criticize you — does more for machine-readable credibility than another round of copy polish. Re-run your ten prompts once a quarter and log the change. That log is your program, even if it lives in a spreadsheet.

If you are the shopper, not the brand

Two things are worth knowing from the other side of the conversation.

First, an assistant’s shortlist reflects what it could verify, not necessarily what is best. A small maker with a thin website can lose to a competitor with better-structured pages. Absence from an answer is weak evidence of quality.

Second, the reasons attached to a recommendation are usually paraphrased from somewhere. When an answer says a product “runs warm” or “holds heat,” it is worth asking where that came from — a spec, a review, or marketing copy. Ask the follow-up question. Assistants are generally happy to tell you what they are drawing on, and the answer tells you how much weight the recommendation deserves.

Third — because the honest version of this list has three items — check the price yourself. Prices and availability change faster than they are read, and a confident sentence about cost is still a sentence about a snapshot.

The short version

The decision moment moved. It used to happen on a results page, then on your product page. Increasingly it happens inside a synthesized answer that a shopper reads once and acts on. Nothing about that requires new marketing magic, but it does require a level of clarity most sites never needed before: readable pages, stated facts, real attributes, credible outside evidence, and someone watching the answers month over month.

How we approach recommendations here, and what we do when we get one wrong, is set out in how we choose, our editorial standards and our corrections policy.