“We Rewrote The Article And The AI Still Names Everyone Else”
Somebody in the meeting says: let’s rewrite that post for AI. So you rewrite it. You add the numbers, you add the sources, you tidy the headings.
A month later you ask the model the same question, and it hands you your competitors again.
I do this too, and the annoying part isn’t the failure. It’s that nothing you did was wrong — the page really is better than it was.
But there’s a study that followed what an AI actually does between your question and its answer, and it suggests the miss is structural rather than a quality problem.
A team at ShanghaiTech University updated the paper in July 2026 (the paper). So let’s look.
How Many Pages Does An AI Read Before It Picks A Name?
An AI with web access doesn’t answer from memory. It searches, opens a page, follows a link, opens another one, and only then commits. An is basically a model that goes and does the looking itself.
In this experiment it got up to five searches and up to five page fetches per question.
So it isn’t reading one page and deciding. It’s walking a route, picking things up, and cross-checking them. And the route, it turns out, is where the contest happens.
They Dropped A Fake Product Into Position Five And Watched
The setup is a little unusual.
The researchers took nine real search results and slotted a fictional product’s page in at a fixed fifth position. Freezing the rank kills the “you just ranked higher” explanation, so what’s left is what happens after the model clicks.
They ran 3,124 query-and-product pairs, using real comparison and recommendation questions.
Then they varied only what sat behind that slot:
- a single page on its own (the baseline)
- that same page polished with three existing methods (content tuning aimed at AI answers)
- six different kinds of page, worded consistently and linked to each other
The team calls the last one TRACE. The six kinds are the official site, reviews, an expert writeup, news, a forum thread, and a social post — with one light hub page out front that points at the other six.
Six Aligned Pages: 67.2%, Against 35.9% For One
So here’s what came out. The share of answers where the model called the product a good fit:
- single page: 35.9%
- six aligned surfaces (TRACE): 67.2%
Two other datasets moved the same way, 56.2% to 71.9% and 59.0% to 73.9% — roughly fifteen points each.
The interesting bit is the control group. Measured inside the same experiment, the three page-level GEO methods landed at 35.9%, 35.9%, and 28.1%. Against a 35.9% baseline that’s nothing, nothing, and worse.
Which is the meeting-room rewrite, plotted.
It Wasn’t The Word Count. It Was The Route.
So what did the work? The team split the conditions again, keeping the same supporting pages and changing only how they sat together:
- scattered, uncoordinated: 75.0%
- wording aligned, linked to each other: 82.8%
- plus the hub page out front: 89.1%
Those levels come from a different condition set than the table above, so don’t line the two up. The thing to read is that the same material climbed step by step on arrangement alone.
Nothing was added. The only thing that changed was the route the model takes while gathering evidence.
And what rose when they added the hub page was, specifically, internal-link traversal. Give the model a directory and it goes deeper; don’t, and it reads one page and calls it.
What This Study Can’t Carry
Some limits, because I don’t want to oversell it.
It’s a preprint, so nobody has peer reviewed it. The experiment ran in a controlled environment rather than the live web. The product was fictional, and GPT-5.1 was the only model tested. The difficulty of winning the rank in the first place isn’t part of the evaluation at all.
The authors also write in their ethics statement that they don’t recommend applying these methods to real systems or the public web, and they aren’t fully releasing the code.
So the thing to take from this isn’t a playbook. It’s the mechanism — how an agent gathers evidence before it answers.
Which reads as “are the surfaces that already talk about us scattered?”, not “let’s manufacture some surfaces.”
Audit The Bundle, Not The Page
Whether an AI names you doesn’t come down to how good any one page is. What seems to matter is whether the surfaces that talk about you exist across kinds, and whether they state the same facts in the same words.
And that isn’t a new demand, when you look at it. The site copy, the case studies, the review-site listing, the conference talk, the community thread — most of these already exist. It’s just that the company name, the category description, and the one-line pitch drift a little on each one. From the model’s side, that drift is what breaks the cross-check. Which also means alignment is cheap upside that’s still sitting there.
Two related pieces run alongside this. The one on quotes and numbers lifting citations 40% is about the inside of a single page; the one on where 150,000 AI citations come from is about the surfaces outside your control. Today’s result says those two stop being separate projects and start being one bundle.
So: list the surfaces that talk about you, by kind, and make the facts match. Then measure whether your name shows up more after the alignment than before — otherwise you’ve tidied things and learned nothing.
Maybe I’m wrong about how much of this survives outside a controlled environment. But aligning one scattered surface is cheaper than building a seventh one, and you find out either way.
Sources
- Hengwei Ye, Jiasheng Mao, Zhenhan Guan, Zheng Tian (ShanghaiTech University), “EcoGEO: Trajectory-Aware Evidence Ecosystems for Web-Enabled LLM Search Agents”, arXiv:2605.12887v2, July 1, 2026, arxiv.org (Starting from the observation that web-enabled AI agents decide recommendations through a multi-step route — searching, following links, cross-checking several pages — the paper moves the unit of optimization from a single page to an “information environment in which several pages support each other”. A fictional target product was injected at a fixed fifth position among nine real search results, and 3,124 query-product pairs were evaluated. The product was recommended in 35.9% of answers with a single page, against 67.2% for TRACE, which aligned the wording and interlinked six surface types: official site, reviews, expert article, news, forum, and social. Two other datasets moved 56.2%→71.9% and 59.0%→73.9%. The three page-level GEO methods did not beat the single page (C-SEO 35.9% / E-GEO 35.9% / AutoGEO 28.1%). A separate ablation under different conditions ran 75.0% uncoordinated → 82.8% coordinated → 89.1% with a hub page added (SafeSearch). Caveats: this is a non-peer-reviewed preprint, tested in a controlled environment on a fictional product with GPT-5.1 alone, and the difficulty of obtaining the rank is not part of the evaluation. The authors state in their ethics statement that they do not recommend applying these methods to real systems or the public web.)
