How to Get Listed on the Sites AI Actually Cites

When a model builds a shortlist, it leans on third-party pages: review platforms, roundups, comparison articles, directories, specialist publications. Your own site establishes what you are. Other people's pages establish that you belong in the set.

Which means the highest-leverage work in AEO usually happens somewhere you do not control — and that is precisely why most teams avoid it.

Step 1: find out which sites decide your category

You do not need to guess, and you should not rely on a generic list of directories.

Ask Perplexity the five questions your buyers would ask. Read the sources it cites beneath each answer. Write down every domain that appears and count how often.

Twenty minutes of this produces a ranked list of the pages that currently decide your category. It is usually shorter than expected — often five to eight domains doing most of the work — and it is specific to you rather than borrowed from a blog post about link building.

A worked example: for a project management tool, that list might come back as G2 (cited in four of five answers), a comparison article on an independent productivity blog (three of five), Capterra (two of five), a Reddit thread in a project-management subreddit (two of five), and one specialist newsletter roundup (one of five). Nothing on that list is a generic "top 100 SaaS blogs" entry pulled from someone else's article — it is the specific five to eight pages actually shaping the answers you tested.

Repeat this quarterly. The sources a model cites shift as new content gets published and old content ages out — a list that was accurate in January can be half wrong by summer, and treating the first pass as permanent is one of the most common ways this work goes stale.

Step 2: sort them by how you get on

The list will contain three kinds of site, and they need different approaches, different timelines, and different amounts of effort.

Self-serve platforms

Review sites and directories where you can create a profile today: G2, Capterra, TrustRadius, Trustpilot, Product Hunt, and whichever niche directories showed up in your Step 1 list. The work is administrative: claim it, complete every field — logo, screenshots, pricing, integrations, category tags — and get enough recent reviews that the profile does not look abandoned.

The most common failure here is not absence but neglect — a claimed profile with four reviews from two years ago and an outdated description. That is worse than no profile, because it exists and describes you badly.

Timing matters more than most teams realise. Ask for a review right after a genuine win — a support ticket resolved well, a renewal, a positive comment in a support thread — rather than in one bulk request months later. Platforms and models both discount clusters of reviews posted the same week; they read as manufactured even when they are not.

Editorial roundups

"Best X for Y" articles by publications and blogs. Inclusion is someone's editorial decision, which means it is a pitch rather than a form — and most pitches fail for the same reason: they read as marketing rather than a fit for one specific article.

The difference is specificity. A good pitch names the exact article, quotes the current entry it thinks is weakest or missing a detail, and states one concrete fact about the product the writer could not have known — a number, an integration, a use case. A generic pitch sent to fifty addresses is recognisable within one sentence and gets the same response: none.

Also worth knowing: many of these get updated. An article from last year that is still being cited may still be edited, and asking to be considered for the next update is an easier yes than asking for a brand-new article to be written.

Community and forum content

Reddit threads, industry forums, question sites. Frequently cited, impossible to control, and actively hostile to anyone who tries to control it.

The only approach that works is genuine participation over time, from people who would be there anyway, answering the specific technical questions in the space for months before ever mentioning the product. When a mention does happen, disclosing the affiliation openly matters — a disclosed mention from an established account tends to survive moderation, while an undisclosed one gets flagged and removed, and a model reading a later snapshot of that thread will not cite content that is no longer there.

Step 3: prioritise by citation frequency, not domain authority

This is where most people go wrong. The instinct is to chase the biggest, most authoritative site on the list.

The better filter is how often a source actually appears in the answers you tested. A niche publication cited in four of your five questions is worth more than a large general site cited in none — regardless of what any third-party authority score says about them.

You are optimising for retrieval, not for a metric no platform reads.

Common mistakes

Spreading effort across thirty low-value directories instead of concentrating on the five to eight that actually got cited. A profile on a directory nobody's answers referenced is wasted time, no matter how easy it was to create.

Sending the same generic pitch to every roundup on the list. It takes the same ten minutes to personalise a pitch as it does to send a form letter, and the response rate difference is not small.

Buying reviews or an undisclosed "featured" placement. Beyond the ethical problem, a page a model treats as advertising carries less weight than one it treats as editorial, and that distinction is usually visible in the markup and the writing itself.

Ignoring a profile that is already claimed but neglected. This is the single most common gap found in Step 1 — not a missing profile, but an abandoned one quietly working against you.

Running the Step 1 test once and treating the list as permanent. The sources cited for a category are not fixed; they move as often as quarterly.

A cadence that actually works

Pitch three to five roundups a week rather than one large batch once a quarter — it is easier to personalise a small number well, and it spreads follow-ups out instead of bunching them.

Follow up once, after seven to ten days, referencing the original pitch rather than restating it. After that, move on. A response rate low enough that ten pitches produce one inclusion is a normal month for this kind of outreach, not a failing one.

What this is not

It is not link building. Models do not count your backlinks when assembling an answer; they read pages that mention you. A link is useful mainly because it usually comes attached to a mention on a page worth reading. An unlinked mention in a well-regarded roundup often beats a link from a directory nobody cites.

It is not paid placement. Some of these sites sell listings, and a paid slot can be genuinely worth it. But a page a model treats as advertising carries less weight than one it treats as editorial — and the distinction is usually visible in the markup.

How to know it is working

Re-run the exact five questions from Step 1 on the same schedule you used to build the original list. Track three things: whether new domains start appearing that were not there before, whether your existing mentions climb from a passing reference to a recommended pick, and whether the answer becomes more consistent across repeated runs of the same question.

Any one of those moving is a sign the work is landing. None of them move quickly — this is the slowest category of AEO work — so measuring monthly rather than weekly avoids mistaking normal noise for failure.

How long it takes

Self-serve profiles can be complete within a week and start influencing answers within a few weeks of that.

Editorial inclusion runs on someone else's publishing calendar, which means months. Community presence runs on years.

This is the slowest work in AEO and usually the most decisive, which is an uncomfortable combination — and the reason it keeps getting postponed in favour of a schema tweak that takes an afternoon and changes less.