Research & TrendsDeep Dive (Vale)
A nautilus behind a counter of six identical steel cheese vaults watches three judges labelled ChatGPT, Grok and Gemini pin first-prize rosettes on three different vaults.

The Top GEO Method on Live Pages Swept All Three Engines for 31 of 80 Brands

In LIVE-GEO, which put GEO-rewritten pages for 80 fictional brands on the live web, the CORE method led on ChatGPT, Grok and Gemini but won all three for only 31 brands. Reversing six links moved its ChatGPT share by 20 points, unsupported statistics still won some top spots, and no apparel was tested.

Neritus Vale

In LIVE-GEO, a test of generative-engine optimisation (GEO) on live web pages, a rewriting method called CORE took the top spot on ChatGPT, Grok and Gemini alike for roughly two-thirds to three-quarters of the fictional brands. By our count of the paper’s brand-by-brand table, it won on all three for 31 of the 80. A GEO lift belongs to the engine, link order and price it was measured on.

Runze Kong of the University of Michigan and colleagues, most at the University of Illinois Urbana-Champaign, published the pages on a startup partner’s public product site and posted the paper on 31 October. Each fictional brand sits in an uncommon category, such as a jacketed cheese vault or an annealing oven, picked “to reduce interference with existing brands and webpages”. Each brand has six product pages built on the same specifications: a template original and one rewrite each from CORE, C-SEO, E-GEO, AgenticGEO and Statistics Addition.

In the headline condition, the test measures reading, not discovery. ChatGPT running GPT-6, Gemini 3.6 Thinking and Grok in Expert mode were handed the product links and told to rank the products using only what the pages said. Five more models ranked the same text without web access. When ChatGPT had to find the products from a brand’s URL instead, CORE’s top-spot share fell from 76.3% to 46.3%.

No clothing was tested. Among the 18 product families, the nearest to apparel is textile and material craft equipment, so any analogy to clothing here is ours.

CORE led on every engine. Table 1 gives each method’s share of brands ranked first, at frozen prices with product links supplied:

Top-1 rate, % of brands ChatGPT Grok Gemini
CORE 76.3 67.5 66.3
Statistics Addition 10.0 11.3 5.0
E-GEO 5.0 7.5 7.5
AgenticGEO 3.8 5.0 8.8
C-SEO 2.5 8.8 2.5
Original, not rewritten 2.5 0.0 10.0

Seven of the paper’s eight evaluators put CORE first most often, though on Qwen2.5-3B its lead over C-SEO was only 1.3 points. The eighth, Claude Sonnet 5.5 reading the text through Anthropic’s API, favoured the untouched original and ranked CORE last on average. E-GEO’s authors read a stable, domain-agnostic pattern as a sign of a “universally effective” GEO strategy. A claim of universality across domains does not carry over to evaluators: here, the best strategy depended on who was reading.

Agreement on the average hid disagreement on the brand. Grok and Gemini, the least aligned pair, named the same winner for 48.8% of brands. For an active-layer thaw-bed planter, Grok chose the Statistics Addition page and Gemini chose CORE’s. Even the closest pair, ChatGPT and Grok, agreed on only about two brands in three.

Link order alone moved ChatGPT’s verdict. With prices equal, CORE took the top spot for 73.8% of brands when its link came last in a fixed list and 93.8% when the list was reversed. The pages were identical. Repeating a run unchanged left the aggregate close to where it was; the paper’s verdict is that visibility is “insensitive to repetition and sensitive to presentation”.

Brand by brand, a repeat was less stable than that aggregate suggests. On ChatGPT, a rerun under the same conditions picked the same top product for only 80.0% of brands, so about one brand in five changed its winner. Some of the disagreement between engines is therefore an engine disagreeing with itself, and the 31-of-80 count comes from a single run.

Randomising is the paper’s remedy. Each brand’s shuffle is drawn once and kept, so the headline figures rest on one order too. On ChatGPT, the shuffled order’s result fell between the two fixed ones. Gemini was another matter: with prices equal, the shuffled order gave CORE the top spot for 82.5% of brands, and either fixed order cut that below 30%. The paper leaves the gap unexplained.

Price moved the result too. The headline figures come from pages priced within 0.2% of each brand’s base price; with every price equal, CORE’s share of Gemini’s top spots was up to 16.2 points higher. The shifts on the other two engines were smaller and did not all run the same way. The paper’s summary, “Values change, but the Top-1 leader remains”, is accurate. A brand lives in the values.

Statistics Addition, a method from the first lab study in our 4 October piece, came second at best by top-spot share on live pages. Its prompt “permits hypothetical data”, and its rewrites made claims like an “18% improvement in heat retention”. In the lab, figures like that moved the engines a great deal. On Gemini it won fewer top spots than the untouched original, yet on average it outranked the original on every engine, and on Grok it had the best mean rank under both fixed orders.

The researchers did the checking. They used an existing verifier, JEV, developed by Deußer and colleagues, which flagged 86 of the 400 rewritten descriptions for numbers the specifications did not support, every Statistics Addition description among them. Because all 80 of those pages were flagged, pushing flagged pages below the rest sets the method’s top-spot rate to zero by construction. The paper concludes that such gains “do not survive the flags”.

No engine saw a flag.

The flags were bookkeeping, kept in a separate file while the pages stayed as written. JEV covered descriptions only, though every method rewrote the review titles and bodies too. Nor does the paper say how much of any lift the reviews supplied. Outside the benchmark, a brand can rewrite its copy but not its customers’ words.

The strongest objection is that the leader barely moved. By our count of Table 1, CORE had the highest top-spot share in every engine run but one. On that reading the scatter is noise around a real advantage, and a brand should buy the method that wins most often. For the thesis to fail, a brand’s own result would have to track that average.

A brand’s own result did not track the average. CORE won all three engines for 31 brands; for the rest, at least one engine crowned another page. Nor was the leader a neutral entrant: Haibo Jin, Huimin Zeng and Haohan Wang, all LIVE-GEO authors, co-wrote CORE. It was also the only method tuned on ChatGPT’s live rankings before its texts were frozen, and it scored highest on ChatGPT. Its optimiser also used GPT-4o rankings, and GPT-4o is among the evaluators, though the paper says the feedback rankings are not counted as evaluation results.

A clothing brand faces the interference LIVE-GEO was built to reduce: real rivals, and whatever each engine already believes about them. If lifts scatter this widely on cheese vaults, a jacket’s lift figure will describe the engine, order and price behind it. A brand can buy that figure, or measure its own pages across engines, link orders and prices; only the second tells it what it bought.

Related Coverage