How Much of the Web Is GEO-Optimized? 9% Overall, 16% of Pages Touched This Year

GEO has had tactics, tools, and think pieces for two years. As of this week it has a number. A preprint posted on 17 August 2026 by Junjie Chu and colleagues at CISPA Helmholtz Center for Information Security, with co-authors at HPE and the University of Waterloo, ran a GEO detector over 10,095 webpages retrieved for 1,000 real search queries and found 8.90% of them were GEO-optimized. Among pages last modified in 2026, the figure was 16.36% — one page in six.

Key takeaways

Everything written about generative engine optimization so far has had to gesture at adoption — vendor counts, agency surveys, anecdote. This paper measured it. The authors sampled 1,000 queries from the ORCAS dataset of real search logs, retrieved candidate pages through both classic Google Search and Gemini-grounded retrieval, and ran their detector over the 10,095 pages they could fetch. It flagged 898.

Two things in that table deserve more attention than the headline.

First, the trend line. Prevalence more than doubles between pages last touched in 2024 and pages last touched in 2026. Whatever else is true, the window in which doing GEO made you unusual is closing: among the freshest content competing for retrieval, one page in six already carries detectable optimization. (The year-by-year figures come from the 1,976 sampled pages carrying modification metadata, so they're the softest numbers in the paper — but the direction is steep enough to survive the caveat.)

Second, the channel gap. Optimized pages are a larger share of what Gemini-grounded retrieval surfaces than of what classic Google Search returns, and the gap widens on recent pages: 18.20% versus 13.52% for 2026-modified content. That is not a controlled experiment proving GEO causes retrieval, but it is exactly what you'd expect to observe if it does — and it's consistent with the pattern we've covered before, where Google rank and AI citation come apart. The pages winning the AI retrieval slot are disproportionately the optimized ones.

The domain breakdown adds texture: 20.37% of sampled Amazon pages flagged, against 4.38% for YouTube and 0% for Wikipedia and major medical sites like WebMD and Cleveland Clinic. GEO adoption is concentrated where commercial intent lives, which is presumably where your prompts live too.

The paper's other contribution is GEOFlagBench: 3,200 pages across 400 queries in health, finance, technology, and travel, with optimized variants produced by eight optimizer families — academic methods like AutoGEO and RAID, "stealthy" variants prompted to hide, and human-performed optimization.

Aggregate results look like a solved problem. A word-level TF-IDF model with logistic regression hits F1 0.880. Pangram, a commercial AI-text detector, hits 0.876. ModernBERT fine-tuned on the benchmark reaches 0.862, and zero-shot LLM judges trail at around 0.77.

Then the authors condition on authorship, and the solved problem comes apart.

The middle columns are the finding. A detector with 0.880 F1 and 0.375 worst-group accuracy isn't detecting optimization — it's detecting AI-sounding prose and getting credit because most benchmark optimization was machine-done. The two failure modes that follow are both practitioner-relevant: honest content that was merely AI-polished gets falsely flagged (Pangram's false-positive rate ran 76.7 percentage points higher on AI-written clean content than human-written), and human-performed GEO walks through. Sparse edits evade too — one light-touch optimizer variant was caught with recall as low as 0.308.

Neither loophole looks durable. The paper's fix, intervention-paired training — training the detector on matched pairs so it learns the intervention rather than the authorship — lifts ModernBERT to 0.944 F1 and 0.883 worst-group accuracy in one move. And a companion result should give anyone running boilerplate optimization pause: detectors could attribute a flagged page to the specific optimizer family that produced it at 0.895 macro F1. The tools don't just leave a mark; they sign their work.

The team also built an auditing agent that inspects the outbound citations on detected GEO pages, grading each source by accountability tier and checking whether the cited URL is accessible and verifiable at all.

Across 6,663 citation occurrences on flagged pages: 68.84% pointed at the weakest source tier, and 69.34% earned the lowest verifiability label. Detected pages surfaced through Gemini carried 8.80 citations each on average, against 4.90 for those from Google Search. Heavily cited, weakly sourced.

This is the empirical shadow of a result we covered last week. The citation wars paper simulated what happens when everyone runs the "add statistics, add citations" playbook and found rewrites accumulating unsupported claims; its proposed defense pays only for claims a source actually supports. GEO-Flag now measures the live-web version: the citations that optimization added are, two-thirds of the time, pointing at nothing a reader could check. The two papers arrive at the same line from opposite directions — one from mechanism design, one from an audit — and the line is verifiability. A citation to a real, accessible, accountable source reads as substance under both. A citation added because checklists say AI engines reward citations reads as a tell.

The usual discounts apply, and a couple of specific ones.

It's a preprint, posted the day before this article, not yet peer-reviewed. The prevalence figure is detector-defined: 8.90% is what a classifier trained on eight known optimizer families flags, which undercounts optimization styles it never saw and inherits whatever authorship bias survives the training fix. The year-over-year trend rests on the minority of pages exposing modification dates. And the retrieval-channel gap is observational — optimized pages co-occurring with Gemini retrieval is not a controlled demonstration that the optimization caused the retrieval.

Most importantly: no platform has announced GEO detection, flagging, or penalties. This paper establishes that the tooling is feasible and publishable, not that it's deployed. The practical read is about incentives — platforms now have a published method for something they already had reasons to want — and about the direction the research keeps pointing, which has been the same direction since the KDD 2024 GEO paper's tactics became the attack model: extractive optimization gets priced out, substantiated content keeps its value.

GEO now has an adoption curve, a detection literature, and an audit trail — the full lifecycle every optimization discipline goes through, compressed into about two years. Nothing here says stop doing answer engine optimization. It says the phase where the web was mostly unoptimized and any structure was an edge is ending on schedule, and the work that survives contact with a detector is the same work that survives contact with a reader: real sources, checkable claims, edits that improve the page.

Elmo is an open-source, self-hosted AI visibility platform that runs your prompt sets across ChatGPT, Claude, Gemini, Perplexity, and Google's AI surfaces and records every answer, mention, and citation — including which domains the engines are actually citing in your category, which is the raw material for the prevalence check above.

For the fundamentals, start with what AI citations are and how to track your brand in AI search. For the arms-race context, read citation wars and our data on citation volatility. For the vocabulary, the AI search glossary.