What AI reads when it answers a buying question

When an assistant answers a buying question it draws on two things: what the model already holds from training, and a small number of pages it fetches at that moment. The fetched set is usually made of four kinds of page: comparison and best of articles, directories and review sites, forum threads, and vendor pages that carry something specific enough to quote.

What is actually happening underneath

Notice the shape of what gets read. A buying question needs a list of options, so the assistant reaches for pages that already are a list of options. That is why the same four kinds keep appearing.

Comparison and best of articles. These are the most quoted sources for buying questions, because they answer in the exact format the assistant needs to produce. One paragraph per tool, already compared.

Directories and review sites. Category pages on review platforms, plus the sector specific directories that matter in your industry. In India these are often local or sector bodies rather than the global review sites, and those local ones are easier to get onto.

Forum and community threads. Reddit, Quora, trade forums, and increasingly long threads on professional networks. They carry something no vendor page can: an opinion the assistant can treat as independent.

Vendor pages, in a narrow role. Your own pages rarely decide who gets named. They confirm facts: what you cover, what you cost, who you are for. A vendor page that only carries adjectives does not even do that.

The number of pages read is small. That is the important part. You are not competing with the whole web, you are competing for a place in a handful of sources.

What makes a page get picked over a near identical one

Two comparison articles on the same subject are not equally likely to be read. Six things separate them, and all six are visible from outside.

It already ranks for the question. The search step happens first, so a page that does not surface never gets considered.

It is in list form. A page structured as one heading per option is easier to pull names out of than the same information written as continuous prose.

It is dated and recent. Buying questions are current by nature, so a page with a visible recent date competes better than an undated one.

It is readable without JavaScript. A page whose body text appears only after scripts run is often fetched as an empty page.

It reads as independent. A third party page carries weight a vendor page does not, which is why your own comparison page cannot do this job for you.

It is specific. Prices, coverage counts, named integrations, stated limits. A page full of adjectives gives the assistant nothing to repeat, so it moves on.

A worked example

We counted this on ourselves rather than guessing at it. Across the whole September 2026 capture on our own site, aiknowsus.com, 24 batches and 72 conversations with Perplexity, a script counted which domains were cited most often.

Two kinds sat at the top: the assistants' own documentation, and the vendors' own websites. A single well known review site appeared far down the list, and no other independent review source came near the top.

Read against the four kinds above, that is a correction worth having. In our category the answers were decided mostly by official pages and by vendors' own pages. A vendor page does more work here than the general advice suggests, and a plan built only around review sites would have been built on the thinnest part of the list.

It also says to count before you spend. That sorted list took a script over our own capture files and one afternoon. Your category will sort differently, and the only way to find out how is to collect your own citations and count them.

What it means for a business

Your content plan has two halves, and most companies only work on one. The first half is pages on your own site that answer one buyer question each. The second half is being present, accurately, on other people's pages. The second half is usually the neglected one and often the faster one.

It also means a page of yours that nobody links to and no crawler fetches is not part of this at all, however good it is.

And it explains why the same work pays twice. A page on a third party site that names you correctly is both a source an assistant can read and a page a human buyer may land on. The two audiences want the same thing here, which is not always true in marketing.

What this does not cover

The four kinds above are what shows up in citation lists for buying questions. They are not a complete account of what an assistant reads, and three gaps are worth stating.

Uncited answers. When no search happened, there is no source list to map, and the answer reflects training data you cannot inspect. Your source map says nothing about those answers.

Sources that are read and not shown. A citation list is what the app chose to display. It is reasonable to assume it is not always the complete set of pages considered, and nobody outside these companies can confirm that either way.

Closed content. Pages behind a login, a paywall or a form are generally not available to these crawlers, so a strong reputation inside a closed community may be invisible here.

One more limit. Your map is built from your questions. Change the questions and the map changes, so it is a map of the ground you chose to look at.

Common questions

Does my blog matter at all, then?

Yes, for two narrow jobs. It is where a page can answer a question no third party page answers well, which is the one place your own site can win the mention outright. And it is where your facts live, so that an assistant checking a claim finds your version.

Is Reddit really used as a source?

Community threads do appear in citation lists for buying questions, because they read as independent opinion. That does not make posting there a tactic. A thread that reads as marketing tends to be removed by the community itself, and the mention is worth nothing without the credibility of the page it sits on.

Should I pay a directory for a premium listing?

Only if that directory actually appears in your citation lists, and only for the presence, not for a promise. Check first. Many companies spend on the famous platforms while the pages actually cited for their category are sector specific and free to join.

How many sources do I need to be on?

There is no threshold worth quoting. The useful target is relative: be present on the pages that get cited repeatedly for your questions, and be present on more of them than the competitor you lose deals to.

Do PDFs and brochures count?

Treat them as a last resort. Text in a PDF or inside an image is far less useful to a crawler than the same text in HTML. If a fact matters, put it on a normal page.

What if no good source exists for my category?

That is the best position to be in, and it is more common in narrow Indian categories than people expect. If no page answers the question well, the page you publish can become the source. Write it as a reference rather than a brochure.

What to do first

• Run ten blind buying questions and collect every cited URL.
• Sort them into the four kinds above. You now have a map of your category's source set.
• Count how many times each page appears. Work on the repeats first.
• Mark the ones that name you. Mark the ones that name your competitors and not you.
• For each unmarked source, decide which of three things it is: a profile you can claim, an author you can contact with real facts, or a page you have to answer with a better one of your own.

Work down that list rather than writing more pages first. Writing is the slower half.