Off-site authority: the GEO lever most guides skip

Eligibility is not selection

It helps to separate two questions an assistant answers in order. First: can this page be read, understood and quoted at all? That is eligibility, and it is what on-site work buys you — a page that is crawlable, indexable, structured, and written so a sentence can be lifted out of it intact. Second: of the many eligible pages, which ones does it actually name? That is selection, and it is decided mostly elsewhere.

This is why sites sometimes fix everything a checklist tells them to and see nothing change on their most valuable questions. Nothing was wrong with the advice. It answered the first question when the site was already losing on the second.

What assistants reach for on a competitive question

Ask an assistant to recommend something and watch what it cites. The pattern is consistent across engines: pages that discuss the category rather than sell within it. In practice that means a handful of recognisable shapes.

• Roundups and "best of" lists — a shape that turns up constantly in the citations for buying questions; they are written to compare rather than to convert.
• Head-to-head comparison pages, including ones published by your competitors.
• Community threads where people describe real use — the assistant treats a specific account of using a thing as evidence in a way it does not treat a product page.
• Reference and encyclopedic entries, which anchor what your brand IS before anything else is decided about it.
• Trade and industry publications, where being covered at all is a signal.

None of those are your website. All of them can be influenced, and none of them can be bought outright without the result reading as bought — to the assistant and to a reader.

What actually moves it

The work is unglamorous and mostly consists of being findable and describable by someone writing about your category.

• Find the pages already cited for your questions, and get legitimately considered for the next revision — most roundup authors accept submissions, and most are working from whatever they could find.
• Make your brand easy to describe correctly: one consistent one-line description everywhere, so an assistant synthesising across sources does not have to reconcile three versions of what you do.
• Publish the comparison yourself, honestly, including where a rival is the better choice. A comparison that only ever wins is not treated as a comparison.
• Show up where your customers actually discuss the problem, as a participant rather than a campaign.
• Claim and keep current the reference entries — the structured directories and profiles that other sources copy from.

What does not work

Being honest about this is more useful than another tactic list.

• Bought links. The pages assistants quote are not chosen by a link graph in the way search rankings were, so the classic tactic transfers badly and carries its usual risk elsewhere.
• Publishing more of your own pages. Ten pages saying you are the best are one source saying it, and the assistant can tell.
• llms.txt. It is worth three points out of a hundred in our own score, and no major engine has committed to reading it — see the honest guide.
• Anything promising a guaranteed position in an AI answer. There is no ranked list to buy into, no index to submit to, and no vendor with a lever on it — including us.

How to know whether it is working

Off-site work pays back slowly and invisibly unless you measure the right thing. Rankings are the wrong instrument: you are not trying to rank, you are trying to be named. Track how often assistants mention you for the questions you care about, and which sources they cite when they do — tracking AI mentions covers the mechanics.

The useful signal is a change in the CITED SOURCES, not just in your own mentions. When a roundup you got into starts appearing in the answers, the cited-source set has changed — watch later runs to see whether your own mention rate changes too.

Frequently asked questions