How AI assistants choose which brands to recommend

Ask an assistant for the best project management tool for a small agency and you get back a paragraph with three or four names in it. Ask a different assistant the same question and you may get a different three. No auction was run. Nobody read your homepage that morning. A sentence was produced, and your brand was either inside it or it was not.

The mechanism behind that sentence is less mysterious than it looks, and it is worth understanding before you spend anything trying to change it.

An answer comes from two places

The first is training. A model has absorbed an enormous amount of text, and brand names sit in that text alongside everything that was said about them. This is not a lookup in a database of vendors. It is a statistical reconstruction of what the written internet tended to say about a category before the training cutoff. A product that appeared repeatedly in sentences shaped like "the best tools for X" is tightly associated with that phrasing. A product nobody wrote about in those terms is not, no matter how good it is.

The second is retrieval. Some assistants search the live web while they answer and write from what comes back. Perplexity does this on nearly every answer and cites as it goes. Gemini leans heavily on current web sources. ChatGPT and Claude search in some situations and answer from training in others, depending on the question and the surface. When retrieval happens, the answer is shaped by whatever pages were returned for a query the model wrote itself, which is rarely word for word the query the user typed.

One honest caveat, because it affects how you read any measurement including ours: answers through the official APIs approximate what the consumer apps say without being identical to them, and the same question asked twice can come back with different wording and sometimes different names. A single answer is an observation, not a constant.

The model is summarising other people's opinions

Here is the reframe that makes the rest of this straightforward. An assistant recommending brands is not evaluating your product. It is compressing what a large number of third party pages already claim about your category and writing a fluent version of that. If a dozen roundup articles, a few review site category pages and a scattering of forum threads all put the same three vendors near the top, the answer puts those three vendors near the top. The model has no independent taste and no way to try your software.

This is why your own website struggles to move it. Your homepage says you are the best option for your audience. So does every competitor homepage, in almost the same words. Across a category, vendor marketing copy is uniformly positive and therefore does almost nothing to separate one brand from another. It is the least discriminating text in the whole corpus.

Your own site still has a job. It supplies the facts that get repeated once somebody else decides to include you: what the product actually does, who it is for, what it costs, which countries and languages it serves, what it connects to. A reviewer writing a listicle lifts those details from your pages, and so does a model that retrieves them. Keep them plain, specific and easy to find. Just do not expect them to get you into the list.

Third party pages decide the shortlist

In practice the sources that shape a category answer fall into a handful of buckets. Review platforms and their category pages. Directories and alternatives sites. Comparison and roundup articles written by publications, agencies and independent bloggers. Forum and community threads where people ask each other what they actually use. Partner documentation and integration pages that mention you in passing.

The asymmetry is stark. One roundup article that names you alongside the incumbents carries more weight in an answer than fifty pages you published yourself, because it is the only kind of page that ranks brands against each other without owning one of them.

Suppose a model is asked for invoicing tools for freelancers in the Netherlands. The pages it retrieves are two English listicles and one Dutch review site. If you appear on none of those and only on your own Dutch homepage, you will not be in the answer, and the size of your actual customer base will not save you. The answer is assembled from the sources, not from the market.

Why the four models disagree

They were trained on different material with different cutoffs, so their inherited picture of your category differs before a single search is run. They retrieve differently, from almost always to rarely. And they write differently: an assistant that hedges and names five options produces a very different shortlist from one that commits to a single winner, which matters a great deal when position inside the answer is what you care about.

That disagreement is information rather than noise. If you appear reliably in the models that answer from training and rarely in the ones that search live, the likely reading is that your historical coverage is decent but there is nothing current for a search to find. The reverse pattern, strong on the searching models and absent from the others, usually means recent coverage that has not yet been absorbed anywhere. Either way, checking one assistant and drawing conclusions about all of them will mislead you.

What actually moves an answer

Start from the sources rather than from a guess. Find which pages are being cited for your category, then work on those specific pages rather than on a general idea of content marketing.

Get listed where your category is catalogued. Review platforms, directories and alternatives sites are unglamorous and they are quoted constantly, because they are structured exactly like the question being asked.

Earn mentions in comparison content written by other people. A "best tools for" article, an "alternatives to" page, a teardown by someone with an audience. You cannot write these yourself and have them count.

Show up where practitioners talk. Community threads are well represented in what models have read. This means answering real questions under a real name, not seeding fake enthusiasm, which readers and moderators spot and which leaves a worse trace than silence.

Be specific about the category you want to be recommended in. Broad claims put you in competition with everyone. A narrow, real description of who you serve matches the way people actually phrase their questions.

Expect different clocks. Pages that get retrieved can change an answer within weeks. What a model absorbed during training changes only when a new model ships.

Where to start

Write down the five questions a buyer in your category would genuinely type into an assistant, with no brand names in them. Ask all four models each question. Record which brands get named, in what order, and which sources are cited underneath. That list of sources is your actual to do list, and it will look nothing like a keyword plan.

Spegla runs that process on ten questions across ChatGPT, Claude, Gemini and Perplexity, and the free check needs no account, but the manual version costs you an afternoon and teaches you more about your category than any dashboard will.