We Asked AI Search Where to Buy Backlinks. It Didn't Mention Us — Here's What It Cited Instead

Most GEO advice tells you to optimise your own website. Add schema, write answer-shaped content, keep your facts consistent. All reasonable — and all of it misses the part that decides whether an AI engine names you.

We tested this on ourselves, and the result was not what we expected.

The test

Two queries through an AI search assistant, a few minutes apart.

The first was branded: our own company name plus "reviews backlink service India". The second was the one that actually matters commercially — a buyer with no brand in mind asking where to buy quality backlinks in India and which service is trustworthy.

Same tool, same day, same topic. One question mentions us; the other doesn't.

What happened on the branded query

The assistant found us and described us accurately. It knew the starter pack is ₹699 for 300 foundation links. It knew the legal entity is Pramila Business Solutions and that a GST invoice comes with every order. It knew about the 12-month free link replacement, and that payment runs through Cashfree.

It even repeated a line from our own site that we had put there deliberately: that we are a new brand without a wall of reviews, and we are not going to fake one.

That is our own copy, retrieved and quoted back. On a branded query, on-site GEO works exactly the way the guides say it does.

What happened on the non-branded query

We were not mentioned. Not in the answer, not in the sources.

The assistant recommended four other services and explained why. It quoted price ranges. It gave advice about editorial placements over volume.

Then we looked at where it got all of that. The sources were not any of those four companies' websites. They were third-party articles — comparison roundups and "best link building services" listicles published by people with nothing to sell in the category.

The engine did not decide who was good by reading vendor homepages. It read what other sites said about vendors, and repeated that.

Why this changes what GEO actually means

There are two different jobs hiding under one label, and most people only do the first.

Being described correctly is an on-site job. When someone already knows your name and asks about you, the engine reads your pages, your schema, your llms.txt. Clean facts in, clean summary out. This is what almost all GEO advice covers, and it works.

Being recommended at all is an off-site job. When someone asks "who should I use", the engine reaches for content that compares options — because a vendor claiming to be the best is worth nothing to it, and a third party ranking ten vendors is worth a lot.

You cannot write your way into the second one from your own domain. We have a well-optimised site, structured data, an llms.txt, and a page targeting this exact category. None of it put us in that answer, because that answer was never sourced from sites like ours.

How do AI search engines choose which brands to recommend?

From what we observed, in roughly this order:

They start from the retrieval corpus for the question — the pages that actually address "which X should I choose". For commercial questions, those are overwhelmingly third-party comparisons, not vendor pages.

They favour sources that name several options. A page listing ten providers with prices is more useful to a summariser than ten pages each claiming to be the best, and it carries less obvious bias.

They repeat specifics. Every recommendation in our test came with a concrete detail attached — a catalogue size, a price floor, a verification method. Vague positioning does not survive summarisation; numbers do.

And they lean on consensus. A brand appearing across several independent roundups gets named more confidently than one appearing in a single article.

What we are doing about it, and what you can copy

The honest short version: get into the comparisons.

That means being genuinely reviewable — real pricing published, a real entity behind the business, verifiable claims — because roundup authors check. It means being present where those authors look. And it means giving them something specific to quote, since a listing that says "affordable and reliable" will be dropped from a summary while one that says "₹699 starter pack, 12-month link replacement, GST invoice" will survive.

What it does not mean is paying for placement in fake "top 10" posts. Those get filtered, they read as promotional to both engines and people, and the whole reason third-party sources carry weight is that they are not bought. Buying them removes the exact property that made them valuable.

Does this mean on-site GEO is a waste of time?

No — it means it is half the work, and it is the half that pays off later.

Branded queries are where buying decisions get finished. Someone hears your name from a colleague, a roundup, or a search result, and then asks an assistant about you directly. If your site is clean, structured and consistent, the answer they get is accurate and useful. If it is not, the engine improvises, and you do not control what it says.

So the on-site work protects you at the bottom of the funnel. The off-site work is what gets you into the top of it. Doing only the first is why a lot of well-optimised sites report that GEO "did nothing" — they optimised for a question nobody was asking about them yet.

How do I check whether AI search mentions my brand?

Run the same two-query test we did, and take the difference seriously.

Ask an assistant about your company by name, and check whether the facts it returns are correct. Wrong prices, an outdated service list or a mangled description are on-site problems you can fix this week.

Then ask the buying question your customers would actually ask — the one with no brand name in it. If you are absent, note which sources the answer cited. That list is your target list. It is usually short, and it is almost never a list of your competitors' own websites.

Repeat both monthly. Answers shift as the underlying content shifts, and you want to notice the direction of travel rather than a single snapshot.

The honest limits of what we found

This was two queries on one tool on one day, in one category. It is a signal, not a study. Different engines weight sources differently, and all of them change without notice.

We also have an obvious interest here: we sell GEO as a service. Which is exactly why we published the query where we lost. A case study where the author is cited by their own method is not worth much; the useful part was finding out we were not.

If you take one thing from this: check where the answer's sources come from, not just whether you are in it. The sources tell you what kind of work will actually move you.