The Service Page Rewrite Trap: What 2026 Research Says

Every AEO consultant gives the same advice: restructure your service pages, add FAQ sections, front-load answers, use question-based headings. The checklist is legitimate in isolation. The problem is how it gets applied -- businesses gut-renovate pages that are getting some citations in pursuit of more, and their citation rates stay flat or drop. The rewrites looked right. The results didn't follow.

Three studies published between April and July 2026 explain why. We've been tracking all three through our research vault, and the convergence is worth a direct explanation.

Getting retrieved is not the same as getting cited

In our October 10, 2026 review of Zhang et al.'s citation absorption framework (arXiv:2604.25707, published in ACL 2026 proceedings), we documented a finding that changes how we explain AI citation to clients: only **15% of retrieved pages actually get cited in the final answer**.

Zhang's team ran a controlled study across 602 prompts, tracking 21,143 citation candidates across ChatGPT, Gemini, and Perplexity. They separated the process into two distinct stages. Citation selection is whether the platform retrieves your page as a candidate. Citation absorption is how much your page's content actually shapes the answer once selected -- whether your language, evidence, and framing appear in what the AI says.

Most AEO guidance -- including nearly everything published about FAQ formatting and heading structure -- targets selection only. It tries to get your page into the retrieval pool. That's a real problem worth solving. But it's only the first half of the problem. If your page gets retrieved and then contributes nothing to the answer, you aren't getting the outcome that matters: your business framing, your service details, your specifics appearing in AI responses.

The finding that matters here: **Q&A formatting as a surface structure does not independently improve absorption**. A FAQ section where every answer is "Yes, we offer this service" or "Contact us for a quote" will earn the citation but shape zero of the answer. The AI selects it, then ignores it. The absorption gain comes from the density of evidence inside the answer, not from the Q&A wrapper around it.

Optimizing for citation aesthetics can damage retrieval

Our October 10, 2026 review also covered a survey by Olivier Martinez (arXiv:2607.14035, submitted July 2026) that synthesizes 45 GEO studies published between 2023 and 2026. The survey's most actionable finding for anyone planning a service page overhaul:

**Content rewrites oriented toward looking citable can degrade the page's retrievability.**

The mechanism: retrieval and absorption have partially opposing requirements. A rewrite that makes a page look more "citation-ready" -- adding FAQ structure, rearranging paragraphs to front-load answers, inserting citation-style formatting -- often strips out the semantic coherence that retrieval systems use to identify and select a page in the first place. The page signals "I want to be cited" while losing the topical depth that earns it a place in the retrieval pool at all.

Martinez's survey proposes an evidence hierarchy that's worth internalizing before touching any client's content:

1. Content that's already been retrieved can be shown to influence citation and answer formation -- this is well-supported. 2. Structural signals amplify citation probability for pages already in the retrieval pool -- moderately supported. 3. Rewrites improve organic discoverability or produce stable long-term citation effects -- **no confirmed causal evidence**.

This hierarchy matters because most AEO advice is operating in zone three while claiming zone one certainty. The surface-level claim ("restructure for citations") hasn't been causally validated for improving retrieval in the first place. What's validated is the middle zone: once a page is already being retrieved, structural signals help. Before retrieval -- they don't substitute.

Document-level architecture beats sentence-level edits

On October 11, 2026, we reviewed FeatGEO (arXiv:2604.19113), published in ACL 2026. This is the most direct academic confirmation of a pattern we've been seeing empirically: **citation behavior is more strongly influenced by document-level content properties than by isolated lexical edits**.

FeatGEO treats a webpage as a set of high-level interpretable properties -- topical completeness, vocabulary alignment, structural coherence, evidence density -- rather than a sequence of words to rewrite. The study tested this approach against token-level optimization baselines (the kind of sentence-by-sentence rewriting that most AEO checklists produce) and found it "substantially outperforms token-level baselines" across three generative engines on the GEO-Bench benchmark.

What the document-level properties actually are:

**Topical completeness.** Does the page fully cover what the query implies, not just the keyword that surfaced it? A page titled "Emergency Plumbing Toronto" that only discusses burst pipe response times is incomplete for queries about weekend availability, water heater emergencies, or drain backups. AI platforms answering those adjacent queries will route past it.

**Vocabulary alignment.** Do the terms in the page match the actual language buyers use when asking an AI? "Comprehensive HVAC maintenance" is professional category language. "Why is my furnace blowing cold air" is buyer incident language. The Discovered Labs β=+0.37 finding (which we documented in May 2026) is the quantitative version of this: vocabulary alignment is the only page-level signal that survived domain fixed-effects controls in their regression.

**Structural coherence.** Modular sections where each H2 contains one complete idea, with direct answers leading each section. Not sentence restructuring -- information architecture. The difference is whether a reader (or an AI retrieval system) can extract a section independently and have it be useful.

**Evidence presence.** Definitions, numerical facts, comparisons. Not assertions. "Over 80% of HVAC emergencies happen outside business hours (ACCA 2025)" is evidence. "HVAC emergencies can happen at any time" is an assertion. Retrieval systems favor extractable evidence over general claims.

What this means if you're planning a page overhaul

The practical implication from these three studies combined: **the sequence matters more than the tactics**.

Before restructuring any service page for AI citations, confirm two things. First, is the page topically relevant to the specific query variants buyers actually use? Not the service category -- the specific incident language. A page about "plumbing services" does not answer "how fast can a plumber get to me for a burst pipe on a Sunday." If it doesn't answer the query, structural optimization returns near zero. Second, does the page rank well enough organically to enter any platform's retrieval pool? ChatGPT cites pages at position 1 in Google at 43.2% -- 3.5x higher than pages beyond position 20. If the page isn't organically visible, it's not in the candidate set, and no amount of content work changes that.

If both conditions are met -- topically relevant, organically retrievable -- then document-level work matters. Rebuild the page architecture: audit topical coverage, align vocabulary to buyer incident language, make every H2 section self-contained and evidence-dense. That's the intervention the research supports.

If either condition isn't met, the right fix isn't a page rewrite. It's Phase 1 infrastructure -- directory presence, entity signals, NAP consistency -- to get the business recognized before any content optimization has a surface to work on.

Platform-specific: Perplexity vs. ChatGPT

The Zhang et al. citation absorption study surfaced one platform divergence worth noting. Perplexity and Google cite more sources per answer -- breadth over depth. ChatGPT cites fewer sources but gives each fetched page higher average absorption influence. The fix plan implication: getting into Perplexity's citation pool at all is the primary win; absorption is distributed. For ChatGPT, citation is more selective, but each citation contributes more to the answer. Evidence-dense pages that survive the 15% retrieval-to-citation threshold matter more on ChatGPT than on Perplexity.

---

We run Signal Check audits specifically to identify whether a business is in the retrieval pool or not -- and which stage of the problem is actually blocking citations. If your service pages have been overhauled and your citation rates haven't moved, the audit usually shows the real bottleneck faster than another round of content editing will.