What Outlets Carry the Most Weight When AI Engines Answer Buying Questions

Independent sites and blogs are where AI engines get their answers in this category. We measured it rather than assuming it, and they account for 79 percent of every citation we recorded. Reddit accounts for 2.7 percent.

That second number is the one worth sitting with, because the figure repeated everywhere is that Reddit drives something like 40 percent of AI citations, and that community platforms account for roughly half of everything an assistant leans on. We kept reading that, and it did not match what we were seeing in our own results, so we went and counted.

What we actually measured

We took 30 real buyer questions, the kind someone types when they are close to spending money, and put every one of them to ChatGPT, Claude, Gemini and Perplexity with live web search turned on. Each question went to each engine twice, which gave us 240 question-and-engine pairs. Between them those answers cited 2,097 sources across 517 distinct domains.

Then we classified every one of those citations by what kind of thing it was. Not by whether we liked it, and not by whether it was the sort of source a strategy deck would recommend. Just what it was.

Four fifths of everything the engines reached for was somebody's article. Not a thread, not a documentation page, not an encyclopedia entry. An article, written by a site that knows its subject, answering the question that was asked.

The Reddit number is not wrong everywhere, it is wrong here

We want to be careful about this, because it would be easy to read our result as a claim that the published research is simply false. It is not. It is measuring a different denominator.

Break our own number down by engine and the picture gets sharper. Reddit reached 4.8 percent on Perplexity, which is the engine most willing to quote a forum. On ChatGPT and Claude it was zero. Not low. Zero. Across 30 questions, run twice, those two engines never once reached for a Reddit thread to answer a buying question in our category.

So if you sell something where the buying conversation genuinely happens in public, on forums, with strangers comparing notes, a Reddit strategy may well pay. Plenty of categories look like that. Ours does not, and the only way to know which one you are in is to run the questions yourself and count what comes back. Somebody else's average is not your measurement. That is the whole lesson, and it cost us less than ten dollars to learn.

Nobody owns this space, and that is the opportunity

Here is the part that surprised us most. Across those 2,097 citations and 517 domains, no single domain accounted for more than 5.3 percent. The most-cited thing we found was a platform's own developer documentation, which is what you would expect for technical questions. The strongest independent sites each took around one percent.

Compare that to a normal search result, where the same handful of enormous sites occupy the first page for every commercial query in a category, and the difference is stark. Search rewards accumulated authority. Citation, at least in our category today, rewards having written the specific answer to the specific question.

Fragmentation like that means the field is open. It also means no single page is going to make you famous, because a one percent share is what a strong independent site looks like here. The wins are small and they accumulate.

The entry price is lower than you think

We also pulled the link profile of every site we saw cited, because we expected the cited sites to be large and established, and we wanted to know how far behind we were.

They were not. One site that earned 19 separate citations across our question set had 18 referring domains. Eighteen. That is a small site by any measure, the sort of link profile a focused operator builds in a few months, and it was being quoted by AI engines answering commercial questions.

Set that against what it takes to rank. For a competitive commercial search term, the weakest site on the first page will typically have well over a thousand referring domains, and the site at the top may have tens of thousands. Those are different orders of magnitude, and they represent different amounts of work and time.

Our own site had nine referring domains when we ran this, and it earned zero citations. We were not close, and we now know roughly how not-close: a handful of links, not a thousand.

Why independent articles win, mechanically

It helps to think about what the engine is actually doing when it answers.

Somebody asks a specific question. The engine runs a search, reads what comes back, and needs to assemble a direct answer it can stand behind. A forum thread is a conversation with no conclusion, five opinions and a tangent, and the engine has to do work to extract anything from it. Documentation is authoritative but narrow, and only answers the question if the question happens to be about that product. An article that sets out to answer exactly the question being asked needs almost no extraction at all. It is already the shape of the answer.

That is why the format matters more than the domain. We saw small sites cited alongside large ones when the small site had written the better answer to that particular question.

It also explains why this is winnable while ranking is not. Ranking is a competition for a fixed number of slots against everyone who wants them. Citation is a question of whether your page is the most useful thing the engine found for one specific query, and there are an effectively unlimited number of specific queries.

What we would do with this, and what we would not

We would write the specific answer rather than the broad guide. A page called "everything you need to know about X" is competing with every other such page, and it answers no question precisely. A page that answers one real question completely is the thing an engine can lift.

We would go after the questions nobody has bothered to answer. The fragmentation says nobody owns this. The absence of a dominant domain means the incumbents are not defending every question, only the obvious ones.

We would not build a strategy on a category average published by somebody else. That is how you end up investing in a channel that accounts for zero percent of your citations on two of the four engines that matter.

And we would be honest that this is a measurement of one category at one moment. Engines change what they reach for. We intend to re-run the same 30 questions periodically, and we keep that question list fixed precisely so the comparison stays meaningful. Changing the questions would reset the baseline and we would learn nothing from the comparison.

What the documentation share is telling you

Platform and vendor documentation took 12.8 percent, which made it the second largest category and much larger than we expected for questions phrased the way a buyer phrases them.

Looking at where it appeared, the pattern was consistent. Documentation got cited when the question had a factual component with a single correct answer. How a particular feature behaves, what a setting does, whether something is supported. The engine reached for the source that owns the fact, which is the right instinct and is exactly what you would want it to do.

This has a practical consequence. There is a class of question you are not going to win, and recognising it early saves money. If the honest answer to a question lives in somebody's official documentation, writing your own version of that answer is unlikely to displace it, and it does not deserve to. Spend the effort on the questions where the answer is judgement rather than fact, because judgement is where an independent voice is the better source and the engine knows it.

The 5.2 percent that came from professional networks, question boards and other forums followed a similar logic. It surfaced where the question was about lived experience, what something is like in practice over months, the kind of thing no vendor will tell you. That is a real category and worth knowing about, but it is small, and it is not where the bulk sits.

How to run this on your own category

You do not need our data. You need yours, and the method is simple enough to describe completely.

Start by writing down 30 questions a real buyer would ask in the weeks before they spend money with you. Not keyword phrases. Questions, in the words a person would use, including the awkward ones about whether the thing actually works and what goes wrong with it. The quality of this list determines the quality of everything downstream, and it is the part people rush.

Put every question to each of the major assistants with web search enabled, and do it more than once, because the same question asked twice does not always return the same sources. Record every citation each answer gives you, along with which engine gave it and which question produced it.

Then classify each cited domain by what it is. Independent site, vendor documentation, community, encyclopedia, retailer. Count the shares. The whole exercise cost us under ten dollars in engine usage and a few hours of setup, and it replaced a strategy assumption we had been carrying for months with a number.

Two rules make the result worth having. Fix the question list before you start and do not change it afterwards, because the moment you edit the questions you lose the ability to compare against your own baseline. And count what the engines actually cited rather than what you hoped they would, including the results that are inconvenient. Ours were inconvenient. We were not cited once, in 240 answers, and knowing that was more useful than any number that flattered us.

Being absent is information too

We appeared in zero of the 240 answers we measured. That is not a pleasant thing to publish, and we thought about whether to leave it out.

We left it in because the zero was the most actionable number in the study. It told us the gap was real rather than theoretical, and when we put it next to the link profiles of the sites that were being cited, it told us roughly how large the gap was. A site with 18 referring domains was in the answers and we were not, which turns a vague worry about authority into a specific, finishable target.

If you run this and find you are absent, that is the same information. It is worth more than an impressions chart going up, because it measures the thing that actually happens at the moment somebody asks for a recommendation.

The trade-off we are not going to hide

Citations are not clicks. A site cited 19 times in an AI answer may see very little traffic from it, because the person reading the answer often has what they needed and never visits. If your business depends on measurable sessions this month, this is the wrong thing to optimise for and we would say so.

What citations buy is different. They put you in the answer at the moment somebody is deciding, and they do it in a context where the engine is effectively vouching for you. That compounds in a way a single visit does not. It is also, right now, unusually cheap to enter, which is the argument for doing it before the entry price rises.

Bottom line

In our category, AI engines answer buying questions almost entirely from independent articles, and hardly at all from Reddit. The space is fragmented enough that no domain holds more than about five percent, and small sites with fewer than twenty referring domains are being cited today. The entry price for citation is a fraction of the entry price for ranking.

If you take one thing from this, take the method rather than our numbers. Run your own buyer questions past the engines, count what they cite, and find out which category you are actually in. Ours did not look anything like the published averages, and yours may not either.