“So Which Of These Thirty Do We Start With?”
Add structured data. Add quotes. Add statistics. I kept writing down things to do about AI search until the table ran to thirty rows.
At last month’s standing meeting the table stayed up on the screen for twenty minutes. And nobody put a finger on row one; we moved to the next item.
But a paper that went back over 45 studies of AI-facing optimisation found two factors that held up across every platform it looked at, and nothing else that did. It’s a preprint, posted in July 2026, and no outside reviewer has been through it yet.
Which AI Search Tactics Hold Up Across The Research?
Two: how relevant a document is to the question being asked, and where that document sits inside the context the model is working from. Those are the only two that survived the conditions changing from one study to the next, after Martinez laid all 45 side by side in a .
GEO (generative engine optimisation) is the work of getting your pages cited or referenced inside an AI’s answer, rather than ranked in a list of links.
The 45 papers were published between 16 November 2023 and 14 July 2026.
So what do those two factors touch, in a job where somebody has to go and edit a page?
Answering The Question Beats Sounding Authoritative
The first factor is the match between the question and the document. Models reach for an explicit match with the query more strongly than for the things a human reads as marks of credibility; that’s how the survey sums up the 45.
But the numbers underneath that reading aren’t the survey’s. They sit in the 2024 paper that coined the word GEO (Aggarwal and colleagues), which built a bench of 10,000 questions, drawn from nine datasets and spread over 25 subject areas, and then measured what each way of rewriting a page did to that page’s visibility in the answer.
Adding quotations moved that visibility up about 44%. Adding statistics and concrete numbers, about 34%. Naming your sources, about 29%. All three are relative gains over a baseline, on a measure that counts how much of the answer your content takes up and weights that by where in the answer it lands.
On the same measure, writing in a more authoritative register came in at about 13%. Keyword stuffing came in at about -8%.
Line those five up and the tactics that make the answer itself more specific sit at the top, while the ones that adjust the tone sit at the bottom. That ordering is what the survey reads as a model picking the match over the authority signal, and I think the ordering carries more weight than the reading laid on top of it.
The person doing that reading is the survey’s author. It isn’t a fresh experiment run alongside the 45, and I’d sooner put that here than let the two factors sound like something that got measured again.
Author’s reading or not, it points at a place I can actually move: the heading, and the paragraph under it. So take the question your reader would actually type, set it in your heading as a plain statement, and finish answering it before the next heading arrives.
Moving A Source Up The Context Beats Rewriting It
The second factor is position inside the context. And pushing a source higher up that context does more than most of the rewrites do.
The is the pile of documents the model has in hand in the moment before it writes. Anything outside that pile stays outside the answer, whatever I do to the prose.
So position here means the order of that pile once it’s been gathered. But who gets put on top is settled by the retrieval step, and that step sits on the other side of the glass from you and me both.
Is there anything on my side of it, then?
Inside one of those documents, where the answer sits is yours to set. Since a page that answers in its first screen and a page that answers five screens down don’t hand over the same fragment, that choice changes what gets pulled out of you.
Although — the survey is measuring documents against each other, not the running order inside a page. I’m slipping one inference in there, and maybe I’m reading across a gap that doesn’t hold; that sentence is the one to argue with.
Get Retrieved First, Then Worry About Placement
General-purpose tricks don’t carry from one platform to the next. A rewrite that worked on one assistant has no confirmation behind it for the next one.
The survey puts the platform point harder than that. Of the 45 studies, not one demonstrated a causal effect on “natural discoverability” that held both across platforms and over time.
After all, many of the gains those studies reported were measured on documents that were already sitting in the candidate pile. So they say something about what happens to a page once it’s in, and nothing at all about what gets a page in.
Which of the two goes first?
Being retrieved does — found by the search step, sitting in the candidate list at all — together with relevance to the question. How your page is placed after retrieval is the second job, and it only pays once the first one is done.
You can redraw the shelf-talker as many times as you like. If the product isn’t on the shelf, nobody’s eyes ever land on it.
What The Forty-Five Were Actually Measuring
The author sorts evidence into five tiers, with field trials that assign conditions at random at the top and synthetic scenarios with a fixed context at the bottom. An is a ruler for how far a finding travels, built from how the study was run rather than from what it found.
But many of the 45 land in that bottom tier, the “synthetic scenarios with a fixed context”. They were measured under conditions that leave out the part where a live search actually pulls documents in.
And this isn’t a new measurement in the first place. It’s a re-reading of 45 papers that already existed (arXiv:2607.14035, posted 15 July 2026). The first page doesn’t say where the author works, either. There’s a contact address and an ORCID (the identifier a researcher gets assigned), and that’s all of it, so weighing this one by who did it isn’t on the table.
The studies it covers are English-language for the most part. If your pages are in English, that’s the case the 45 cover; for any other market you sell into, whether the same two factors hold up is a question nobody has put a number on.
Which leaves the checking to you, and on my own pages, to me.
Split The Checklist Into Relevance, Position, And Everything Else
Draw three boxes beside your AI search table. Rows that work on relevance to the question. Rows that work on where your page sits once it’s been retrieved. Rows that are neither.
But the third box doesn’t get thrown out. Nothing across the 45 showed those rows were useless — a consistent effect just didn’t turn up for them, and I’d rather demote a row than claim a result the research never produced.
Relevance goes first.
The position box opens once the retrieval side is through.
How far individual tactics move citations is in The Trick To Getting Cited By AI Is Boring: Add Numbers And Sources, and why a search rank and an AI citation are two different things is in Can’t Win At Search, So You’re Doomed In AI? The Long Tail Says No.
My thirty came apart into nine and twenty-one. Twenty minutes of silence in that meeting, and the same table turns out to be pointable once it’s nine rows long.
Sources
- Martinez, O., “Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026)”, arXiv:2607.14035, 2026-07-15, https://arxiv.org/abs/2607.14035
- Aggarwal, P. et al., “GEO: Generative Engine Optimization”, 2024, https://arxiv.org/abs/2311.09735
