Title and meta rewrites gained 22 percent of retrievals in a GEO benchmark

What the benchmark ran, and on what

The corpus was built rather than borrowed. The authors sampled 300 queries from each of nine established information retrieval datasets, giving 2,700 unique queries spanning web search, general question answering, multi-hop question answering, biomedical, community question answering, finance, debate, shopping and academic material. For each query they retrieved up to 100 results through the Google Custom Search API and crawled every URL. After discarding malformed documents the corpus holds 171,003 unique web pages, and each query is associated with 63 candidate documents on average.

The part that makes the corpus unusual is what survived the crawl. Earlier benchmarks in this area index body text and throw the rest away. This one keeps five separate fields per document: the title, the meta description, the headings from h1 to h6, the schema markup as JSON-LD, and the body text. The paper says those four structural fields were chosen because they are the ones public search engine guidelines consistently emphasise, and it indexes them as separate components rather than merging them into one string, on the stated grounds that production retrieval systems score fields independently and combine them later.

The engine itself is three stages. BM25 does first-stage retrieval, with a separate index per field and reciprocal rank fusion across them. Qwen3-Reranker-4B reranks the candidates as a cross-encoder scoring each query and passage pair. GPT-5-mini generates the answer from the top reranked candidates and is prompted to cite its sources. That is a conventional retrieval-augmented generation stack, and naming its parts matters more than it might seem, because it is the reason every figure below is a statement about this pipeline and not about any product a reader uses.

The measurement procedure is a before and after on one document. For each query the pipeline runs once, a target document is picked at random from those that reached the generation stage, and that becomes the baseline. The target is then rewritten by one strategy, reindexed, and the same query is run again. The paper tracks whether the document still appears in the top k at each stage, using k of 100 at retrieval and k of 10 at reranking, and how far its rank moved. Because targets are drawn from the top ten after reranking, the reranking hit rate at k of 10 is 1.00 in the baseline by construction, which is worth holding on to when reading the reranking numbers later.

This is the end-to-end design that July's survey said the field was mostly not attempting. That survey put crawling and indexing at stage two of a seven-stage pipeline and reported that far fewer studies look there. SAGEO Arena is one of the studies that does, and what it finds at those earlier stages is the reason to read it rather than the citation numbers at the end. Nothing here observes an AI crawler in the wild: the documents were fetched once, by the authors, and then held still.

Why the GEO body rewrites lost ground at retrieval

This is the half already summarised on this site, so the point of restating it is the detail the summary flattened, which is the spread across strategies and the mechanism the authors give for it. Eight of the ten strategies are adapted from the original GEO paper: authoritative tone, cite sources, fluency, quotation, easy language, statistics, technical terms and unique words. The ninth is All-in-One, which applies all eight at once and then adds structural formatting such as bolding and layout changes. The tenth is AutoGEO, which rewrites a document using preference rules learned from watching generative engines pick citations. Rewriting only the body text with any of them left the document worse off at retrieval, and the average across all ten is a hit rate of 0.53 against a 0.58 baseline, with the average rank moving 4.54 positions in the wrong direction.

The spread is wider than the average. The two vocabulary strategies did the most damage among the eight: technical terms took the hit rate to 0.50, a fall of 14 percent, with an average rank change of minus 6.23, and unique words took it to 0.53 with minus 3.47. The paper's explanation is mechanical rather than mysterious. A BM25 retriever scores lexical overlap, user queries use ordinary vocabulary, and a rewrite that replaces eating with alimentary routines or sleeping with somnolence removes exactly the terms the query is going to contain. AutoGEO was the worst by a distance, taking the hit rate to 0.37 and the average rank change to minus 22.35, which the authors attribute to its habit of expanding documents at length and diluting keyword density.

At the generation stage, where these strategies were designed to work and where they are usually evaluated, the effect was small and negative: 0.47 against a 0.50 baseline, a fall of 6 percent averaged over the ten. That is consistent in direction with what a controlled two-source citation study reported in May, where the two formatting factors sat among the seven with no consistent effect across six models, and where the factors that did move the first citation were mostly not presentational. Two different designs, two different corpora, and the same shrug about presentation.

What is new here is the compounding. A strategy that costs a document three or four positions at retrieval is not trading rank for citations, because a document that misses the retrieval cutoff never reaches the generator to be cited or not cited. The paper puts it plainly: only top-ranked documents are passed to the generator, and a document that falls below the threshold is invisible at generation regardless of content quality. A rewrite evaluated on a fixed candidate list cannot see that cost, which is what the fixed candidate list was hiding.

There is a separate and more uncomfortable finding about repeated rewriting that this benchmark does not test, and it is worth reading alongside: an automated optimization loop degraded the document over five consecutive rounds, accumulating unsupported claims in the worked example. Google's own guidance, for its part, names five tactics it says do not help. The three findings do not agree on method or scope, and they point the same way about rewriting prose to be found.

Which fields the 22 percent came from

Run the same ten strategies against the structural fields instead, leaving the body text alone, and the retrieval numbers invert. The average hit rate goes to 0.71, a relative gain of 22 percent, and the average rank change goes to plus 2.72. Eight of the ten strategies gained at retrieval in this scope, with fluency at 0.75 and plus 6.62, statistics at 0.75 and plus 6.03, cite sources at 0.74 and plus 6.05, quotation at 0.74 and plus 5.47 and easy language at 0.74 and plus 5.15. Only AutoGEO, which rewrites at length in any scope, still finished below the baseline on rank.

The fields in question are unglamorous. Title, meta description, headings and structured data as JSON-LD: the same four things that on-page checklists have listed for two decades. The paper's account of why they moved the number is again mechanical. Structural fields are short and dense with the entities a query is likely to name, so rewriting them raises lexical overlap with the query in a way that rewriting a long body text does not. The worked examples in the paper are small and concrete. A verbose meta description becomes a summary carrying specific facts, given as 1983, comedy, Rotten Tomatoes, grossed $61M. A title reading Panel Clarifies Advice becomes IOM Panel Clarifies Vitamin D Guidance.

The counterintuitive row is the third scope. Rewriting both the body text and the structural fields gave 0.67 and a rank change of minus 0.56, which is worse than rewriting the structural fields alone on both measures. The negative retrieval effect of the body rewrite partially cancels the structural gain. Whatever else that says, it says the two edits are not additive, and it says the common advice to do everything at once has a measured cost in this pipeline.

None of this is a measurement of any site a reader owns, but the fields are ones Lantad has counted on real pages, and those counts have their own dates. Across five pages this scanner captured and stored on 15 July 2026, we published on 18 August 2026 that four JSON-LD blocks held 102 typed nodes, with one page accounting for 96 of them and 39 of those describing a navigation menu. We published on 24 August 2026 that the same five pages carried 85 meta elements between them, of which two were robots meta tags carrying nothing but an image preview size. And we published on 17 August 2026 that the raw HTML held 245 heading elements while our extractor recorded 166, because headings inside nav, header, footer or aside are discarded. Those are our measurements of five stored files. They say the fields exist and what is in them. They say nothing about retrieval, because we have not measured retrieval.

Every scope lost ground at reranking

The stage the paper calls a persistent bottleneck is the middle one, and it is the finding least likely to appear in a summary of this work, because it is bad news for every column at once. Body text rewrites took the reranking hit rate to 0.84. Structural rewrites took it to 0.83. Doing both took it to 0.75. All three are falls from a baseline of 1.00, and that baseline is 1.00 by construction, since every target document was drawn from the top ten after reranking before anything was changed. In other words, whatever was done to these documents, some of them stopped surviving a stage they had already survived.

The magnitudes are small in absolute terms. The average rank drop stays inside one position across most settings. The paper's argument is that small is not the same as harmless when a pipeline imposes a cutoff, and it gives the number that makes the point: 5.8 percent of target documents dropped from rank 10 to rank 11 during reranking, which in this configuration means out of the generator's input entirely. A document at rank 11 is not a document that ranks slightly worse. It is a document that is not in the answer.

The case study on which documents gained and lost rank produced two patterns worth repeating, because they are the closest thing in the paper to actionable advice that does not depend on its specific model choices. The reranker rewarded additions that spoke more directly to the query's informational need, and penalised additions that widened the document's scope beyond it. And it rewarded putting the answer early: restructuring that pushed the answer into later paragraphs cost rank sharply, even when the answer itself was untouched and still present. The authors note that optimization here happens without knowing the incoming query, which is the real constraint a page author works under too.

There is a reason to be careful about carrying any of this over to what a commercial product reports. The rank a document holds inside one pipeline on one run is not a stable quantity, and Lantad has published its own version of that problem: our prompt tracking needed seven runs per prompt before the measurement settled, and we have argued separately that a citation count is not the same as influence on an answer. Any figure describing AI visibility is a distribution, and a single-run rank change is a sample from one.

Structural fields got documents retrieved, then went unquoted

The paper runs one more experiment that explains why the two scopes behave so differently, and it is the one worth the most to anyone deciding where to spend an afternoon. The authors prompted the generator to return the exact quote behind each citation it made, then fuzzy matched each quoted span back to a region of the source document, and plotted where citations came from. The result is that the overwhelming majority of citations originate in body text, and structural fields are quoted far less often despite their much stronger contribution to retrieval.

Put the two findings together and you get a division of labour rather than a winner. Structural fields decide whether the document is in the room. Body text decides whether it is quoted once it is there. The paper's own summary is that the two serve complementary roles and that effective optimization has to address both, which is a less quotable conclusion than either half on its own, and is the one the evidence supports.

The backbone comparison is the sharpest illustration of the split. Running the All-in-One strategy through three different rewriting models, LLaMA-3.3-70B produced short keyword-dense text averaging 296 words against 782 for GPT-5-mini, and won the retrieval stage most often, with a 42.2 percent win rate against 38.0 percent for GPT-5-mini. Then it collapsed. At reranking GPT-5-mini won 63.1 percent of the time against 13.0 percent for LLaMA-3.3-70B, and at generation the gap widened to 73.2 percent against 7.0 percent. A document engineered to match queries lexically was retrieved and then discarded by everything downstream that reads for substance.

One domain result belongs here as a caution against generalising. Across the nine domains, shopping was the only one where every optimization method reduced citation likelihood. The paper's reading is that product pages are already organised for their buyers, and that the queries in that domain are casual and recommendation-seeking, so rewriting the page pushes it away from the register the generator is looking for. If you sell things, that is the single most relevant sentence in the paper, and it points the opposite way to most advice aimed at you.

The idea that one blob of markup can score well on one reading and badly on another is not new here either. Lantad reads the same JSON-LD twice with two different traversals, and on our stored Wikipedia capture one reading returns a perfect 1.0 and the other returns 0. Similarly, a page's entity confidence signals often sit in exactly the regions that a generator rarely quotes from, and our count of where visible text sits found that more than half the text blocks on five real pages were navigation. The regions of a document are not interchangeable and they are not equally weighted, which is what this benchmark measures directly for the first time.

What this does not say about your site

The benchmark is a research pipeline. BM25 is a lexical scorer from 1994 and it is the reason the structural fields helped as much as they did, because short dense fields are exactly what term-frequency scoring rewards. A commercial answer engine may use a dense retriever, a hybrid, a live search index, or a cached one, and the paper does not claim otherwise. Nothing in it measures ChatGPT, Google AI Overviews, Perplexity or Claude, and no figure above should be quoted as though it did.

Google's own position points the other way and deserves stating in full rather than being explained away. Its guide to optimizing for generative AI features, last updated on 10 July 2026, says that structured data is not required for generative AI search and that there is no special schema.org markup you need to add, while recommending it anyway as part of ordinary search work. It also says you do not need to create new machine-readable files, AI text files, markup or Markdown to appear in Google Search including its generative capabilities. Those two statements and this benchmark are about different systems, and both can be true at once. A reader who wants a rule out of this should take the weaker one: structural fields cost little, are useful for other reasons, and were the only scope in this experiment that improved first-stage retrieval.

What Lantad measures is narrower than either, and worth being exact about. We measure whether the fields are readable at all in the response a crawler receives, which is a prior question to whether they help. Our composite weights prose parity at 50 points of 100, access at 25, structure at 15 and schema at 10, and those weights are settings chosen by us rather than findings from anything. The scan reads what arrives in the initial HTTP response, and you can see the same view for a page you own with what GPTBot sees. Our methodology states the boundaries of that, and our research page publishes what the scan population supports.

That prior question is not academic on a modern stack, and it is the practical carry-over from this paper. Title, meta description, headings and JSON-LD all live in the head or the early body of a document, which is precisely where a client-rendered application often has nothing until hydration runs. We have measured how that ends: two stored fixtures produced an identical composite of 34.4 and an identical F, one shipping zero words in its initial HTML and the other shipping all 485 of them inside a JSON payload. If the four fields this benchmark found most useful are injected by a framework after the fact, they are not in the document that a first-stage retriever indexes. That is a fixable defect on most stacks, and the per-framework detail is in our guides for Next.js and React.

So the honest summary is short. On one research pipeline, over 171,003 documents and 2,700 queries, rewriting prose for citations cost retrieval, rewriting the four structural fields gained it, doing both gained less than doing the second alone, and every scope lost ground at reranking. Whether the same holds inside any commercial engine is unmeasured, by these authors and by us. What is measurable today, on your own site, is whether those four fields are in the response at all.