What AI assistants recommend, and why it is not you

AI assistant recommendations for software come from third-party pages, not from the vendor. In March 2026 we put 200 purchase-intent questions across five SaaS categories to ChatGPT, Perplexity and Gemini and logged every source they cited. The pages that earned the recommendation had three things in common: they were written by someone other than the vendor, they answered the exact question in the first paragraph with a named pick and a reason, and they carried a visible recent date. Vendor pages were cited too, but for facts like price and limits, almost never for the pick itself. This is a small sample from one month, so treat it as a pattern to test against your own category, not as a statistic.

How we ran the sample

We wrote 40 questions in each of five categories: cold email warmup, screen recording, invoicing for freelancers, uptime monitoring and form builders. Every question was the kind a buyer types when they are about to choose, with a constraint attached: "best email warmup tool for a 3-person agency", "screen recorder that exports to GIF under 5 MB", "invoicing app for a freelancer who bills in two currencies". We asked each question once in each of the three assistants with web access on, in a fresh session, and copied every cited URL into a sheet with the date, the assistant, the question, and whether the page was a vendor page, a third-party page, or a community thread.

Caveats first, because they matter more than the findings. One month, one set of questions, one person writing them. Assistants change their retrieval and ranking often, and answers vary between sessions. We did not measure how many people ask these questions, and we have no way to; nobody outside the assistant vendors does. What we have is a log of which pages got cited when a specific question was asked, and a pattern that was consistent enough across 200 questions and three products to be worth writing down.

What AI assistant recommendations have in common

Three properties showed up again and again in the pages that carried the actual recommendation, as opposed to the pages cited for a fact.

1. Third-party authorship. The recommending page was a newsletter, a blog, a review site, a YouTube video description, or a Reddit thread. When a vendor page appeared in the citation list it was almost always next to a price or a feature claim, not next to the sentence that said "use X".
2. The exact question answered in the first paragraph, with a named pick and a reason. Pages that opened with a definition of the category or a history of the problem were cited less than pages that opened with "for a three-person agency, use X because it does Y".
3. A visible, recent date. Where two pages covered the same ground, the one with a 2025 or 2026 date on it was cited more often than an older, longer article, even when the older article ranked higher in a normal search.

Why vendor pages get cited for facts, not picks

This is not a mystery and it is not a bias against vendors. An assistant answering "which should I use" is being asked for a judgement, and the retrieval step looks for pages where a judgement already exists. A vendor page does not contain a judgement about alternatives; it contains claims about itself. So the assistant takes the claim it can verify from the vendor (the price, the limit, the integration) and takes the judgement from someone who compared. The assistants' own documentation describes answers as grounded in retrieved sources with citations shown to the user [1][2], and a source for "best" is, by construction, someone who looked at more than one option.

There are two practical consequences. First, your own site cannot write its way into the recommendation, no matter how good the page is. Second, your own site must still be good, because when the assistant pulls your price or your limits it will pull them from your page, and if the page is vague or out of date the answer will be too. A pricing page that says "contact us", a terms page with no numbers, or a feature list that has not been updated since a rename will all show up in answers, verbatim and wrong.

Recency: why the date on the page matters

The strongest single pattern in the log was the date. Buyer questions are time-sensitive by nature: prices change, products add features, tools shut down. The assistants appeared to prefer pages that were visibly recent, and "visibly" is doing work in that sentence. A page updated last month with no date on it looked older to the assistant than a page with "Updated February 2026" in the byline. Google's own documentation on publication dates recommends showing a clear date and keeping it accurate, and the same discipline helps here [3].

Recency also explains why long, authoritative articles lost to shorter, newer ones. A 4,000-word guide from 2023 that ranks first in search was cited less than a 900-word post from January 2026 that answered the same question. This is uncomfortable for anyone who invested in evergreen content, but it is also an opening: the pages winning AI assistant recommendations right now are not expensive to produce, and most of them were written by one person with direct experience.

What this means for a brand

Put the three properties together and the conclusion is uncomfortable but simple. To be recommended by an AI assistant, someone else has to say it, on a page that is specific, recent and answers a real question. That someone is a creator, a reviewer, a newsletter writer or a community member. It is exactly the set of people an affiliate program already pays, and exactly the set of people most affiliate programs pay only for clicks and sales, never for the page itself.

Here is a worked example from the sample. For "invoicing app for a freelancer who bills in two currencies", one newsletter issue was cited by all three assistants. It was about 800 words, dated, written by a freelancer who had used three tools, and it named one with a reason: the multi-currency invoice did not need a separate template. The author had a partner link in the issue. The vendor named in that issue got the recommendation in three assistants without writing a word, and the author got paid on the signups. That is the whole mechanism, and it is already running for whoever is on the right side of it.

If you are the vendor who was not named, you have two choices. You can hope a reviewer picks you next time, or you can find the people who write pages like that one and give them a reason to write one about you. The second choice is an affiliate program with a content brief attached: here is the question buyers are asking, here is what we would like you to test honestly, here is a bounty for the delivered piece and a commission on whatever it brings in.

A procedure: find the questions you are losing

You can run a version of our sample for your own category in an afternoon. It is tedious but not hard, and the output is a list of questions you can hand to partners.

1. Write 30 to 50 buyer questions with a constraint in each: team size, budget, integration, format, region. Avoid "what is the best X"; use "best X for someone who needs Y".
2. Ask each question in ChatGPT, Perplexity and Gemini with web access on, in fresh sessions. Record every cited URL and the sentence it supports.
3. For each question, mark who got the pick: you, a named competitor, or nobody clear. Mark whether your own page was cited for a fact, and whether that fact was correct.
4. Sort by two things: questions where a competitor got the pick, and questions where nobody did. The second group is the cheapest to win because there is no incumbent page to displace.
5. For each priority question, write a one-paragraph brief: the question, what an honest answer would need to test, and what a good page looks like (first-paragraph answer, named pick, date, specifics).
6. Fix every wrong fact the assistants pulled from your own site. This costs nothing and takes an hour.

The sister product InsightWonder does the first four steps continuously per brand: it finds buyer questions where assistants cite competitors, nobody, or you, and keeps the list current. It does not estimate how many people ask each question, because nobody can measure that honestly from outside, and we would rather say so than invent a volume column. Later this year RelayWonder will take those gaps and turn them into content campaigns a partner can accept, with the delivered URL tracked for clicks, sales and AI citations.

Make the page easy to cite

Whether the page is written by a partner or by you (for the facts), the same structure makes it more likely to be used. None of this is a trick; it is the same advice Google gives for helpful content, applied to a reader that is a model [4].

• Answer in the first paragraph. State the pick and the reason in two sentences before any background.
• Use the buyer's constraint in the heading and the first sentence, in the words a buyer would use.
• Show a date, and update it when the content changes. A byline with a person's name helps too.
• Give specifics that can be verified: a price, a limit, a format, a screenshot described in words.
• Add a short FAQ with the two or three follow-up questions a buyer would ask next, marked up with FAQPage structured data where appropriate [5][6].
• On your own site, keep pricing, limits and program terms in plain numbers on pages that do not require a login. Those are the pages assistants will quote.

We will rerun the sample later in the year with the same questions and publish what changed. If the pattern holds, the brands winning AI assistant recommendations in 2026 will be the ones with the most partners writing specific, dated pages, not the ones with the biggest content budget.