The guideChapter 2

What works in AI search, graded by evidence

Published Last reviewed

In short

  • Do the work that makes your pages accessible and their information dependable.
  • Treat proposed citation gains as hypotheses that need evidence from the relevant engine and task.
  • Don't buy schema, FAQ blocks, llms.txt, added statistics or artificial chunking as established citation levers.

A proposal can include useful work and still overstate what it will achieve. Fixing a blocked page has a clear acceptance check. Promising recommendations after adding schema requires a different kind of evidence.

Our verdicts below are Red-engage's recommendations. The sources establish the documented requirements or research findings behind them. A "Do" verdict doesn't mean a controlled experiment proved a citation increase.

What each verdict means

Do covers foundations and editorial work we recommend for eligibility or accuracy. Test covers ideas worth a limited experiment when they answer a real business need.

Don't buy it as a lever rejects the citation promise. The underlying practice may still help readers or support another search feature. Avoid covers manipulation and misleading presentation.

Read evidence at its actual scope. Platform documentation describes that platform. A controlled study tests its own setup. An observational study can reveal an association without establishing what caused it.

The practice verdicts

Each row separates our buying decision from what its sources support. "Editorial judgment" means we recommend the work without claiming a demonstrated citation gain.

PracticeVerdictEvidence and useful purpose
Check search crawler access through robots.txt and the CDNDoGoogle documents both layers. OpenAI distinguishes search access from training access. This addresses eligibility. 1 2
Check indexing and snippet eligibility for important pagesDoRequired for supporting links in Google's AI features. Eligibility leaves selection uncertain. 1
Review Search Console's Search generative AI controlDoGoogle documents a separate inclusion choice, including inherited settings. 11
Put essential answers in accessible page textDoGoogle asks for important information in text. Our preference for initial HTML simplifies verification. 1
Maintain accurate product, pricing and public help pagesDoEditorial judgment applying Google's reliability guidance to factual questions about the business. 3
Publish a visible, dated brand facts pageDoEditorial judgment for maintaining verifiable facts, grounded in Google's guidance on accuracy. No citation gain is established here. 3
Make substantive corrections and document human reviewDoGoogle requires checking AI-generated material, including metadata. Its content guidance rejects superficial date changes. 3 4
Maintain relevant ecommerce feedsDo, for ecommerceGoogle documents Merchant Center as a source of product information for Search, including AI features. 5
Include platform reports in measurementDoSearch Console and Bing document AI performance reports. Bing citations don't establish a recommendation. 12 13
Publish original research or a useful toolTestGoogle encourages original value. Citation impact remains the experiment, with usefulness judged separately. 3
Publish a supplier comparison on your own siteTest, with editorial approvalOur judgment is to require substantiated criteria and useful analysis. Google's quality guidance doesn't validate a citation tactic. 3
Improve agents' ability to complete a real taskTestGoogle describes agents interacting with websites. Test task completion before claiming any visibility effect. 5
Add a separate, publicly readable AI knowledge sectionTest, with editorial approvalOur limit is a useful, small experiment without a volume target. Reader value must justify it. Hidden manipulative text is excluded. 3 10
Add schema to increase AI citationsDon't buy it as a leverSchema helps describe content and supports rich-result eligibility. Observational citation evidence has substantial limits, discussed below. 6 9
Add FAQ blocks or tables to increase citationsDon't buy it as a leverOne observational preprint challenges FAQ-based claims, with limits discussed below. It doesn't establish a table effect. Choose either format for the reader's task. 9
Add statistics or expert quotations to get citedDon't buy it as a leverEarly controlled findings didn't transfer in a later controlled replication. Keep useful evidence for its substance. 7 8
Add llms.txt, Markdown copies or a special AI layerDon't buy it as a leverGoogle explicitly rejects these as visibility improvements for Search. This statement concerns Google. 5
Split prose into artificial chunksDon't buy it as a leverGoogle says tiny content fragments aren't required. Choose section lengths for readers. 5
Mass-produce near-duplicate pages for question variantsAvoidScaled content intended to manipulate Search violates Google's spam policies. 10
Change dates without changing the substanceAvoidGoogle flags this as a search-first practice. A revision date should describe actual work. 3
Hide promotional text, inject instructions or deceive crawlersAvoidGoogle's policies cover hidden manipulation and cloaking, including attempts to manipulate its AI answers. 10

Why the research doesn't justify a guaranteed recipe

The original GEO paper reported gains from interventions including statistics and quotations. Its controlled setup supplied selected sources to an answer generator. 7

The Perplexity experiment also supplied source text as uploaded files. Answers were constrained to those files, as Appendix C.1 explains. That tests behavior after sources are supplied. It doesn't show that editing a website makes Perplexity discover it. 7

In a September 2026 preprint, Bajemon and Rochet retested quotation, statistics and source-citation edits. They didn't reproduce positive citation gains under their controlled protocol. The authors are affiliated with TW3 Partners, identified as Citead. 8

Their experiment also imposed sources and citation instructions. It cannot establish that these edits never help anywhere. We treat it as a reason to reject a universal sales claim. 8

Moore and Dunne's September 2026 preprint studied citations across B2B SaaS workspaces. The authors report that apparent benefits from FAQ blocks and structured data weakened or reversed once domains were taken into account. 9

The preprint describes its analysis sample inconsistently between sections 3.3 and 5.2. That limits how much weight these findings can carry. 9

The authors are affiliated with Discovered Labs. Their study was observational, and schema results covered third-party pages because extraction failed on some brand-controlled pages. Its findings cannot establish a causal result for a new ecommerce site. 9

Our conclusion is narrower than "format never matters." We won't sell a formatting change as an established citation lever on this evidence.

Keep the useful work

Schema can help Google understand page content and make eligible pages available for rich results. Keep it consistent with what people can read. 6

A pricing table can help a buyer compare plans. A short answer can resolve a recurring support question. Those are editorial reasons to choose a format. Ask the customer-facing team whether the change actually makes its job easier.

The same applies to evidence. A relevant measurement with a clear method deserves inclusion. A quotation should add expertise the paragraph needs. We don't recommend decorating a page with either to satisfy a citation quota.

For factual pages, define success before discussing discovery. If an answer misstates a cancellation policy, the first task is to make the correct policy clear. Save the answer and verify the replacement information with the person responsible for it.

Google doesn't require llms.txt or artificial chunking for its generative Search features. Its guidance explicitly says those special files don't improve Google visibility. 5

That leaves room for a separate operational purpose, such as documentation consumed by a known integration. Require the proposal to name that purpose. Don't quietly relabel it a citation strategy.

Run tests that can change a decision

We recommend writing a short experiment brief before editing. Name the buying question and the information gap. State the change you propose and what observation would justify keeping it.

For example, test a clearer explanation of a product limitation on an existing page. Preserve the original version. Keep the prompt set stable and compare the same engines under comparable conditions.

Track accuracy separately from linked citations. Inspect recommendations independently. Where practical, leave comparable pages unchanged so broad shifts are easier to notice. A before-and-after difference alone should remain an observation.

Use platform reports within their stated scope. Bing's report counts source citations across supported experiences, without establishing a page's role in an individual answer. 13

Agree a review point based on available observations. A sparse result can remain inconclusive. It doesn't justify increasing page volume to make the chart move.

Carry the protocol into measuring AI visibility. For access issues, use technical foundations. To assess a provider's scope, use the agency proposal review.

Buy work whose purpose survives inspection. Keep the promised result within what the evidence can support.

Sources

  1. Google Search Central, AI features and your websiteAccessed Back to the text, first mention of source 1
  2. OpenAI, Overview of OpenAI CrawlersAccessed Back to the text, first mention of source 2
  3. Google Search Central, Creating helpful, reliable, people-first contentAccessed Back to the text, first mention of source 3
  4. Google Search Central, Guidance on using generative AI contentAccessed Back to the text, first mention of source 4
  5. Google Search Central, Optimizing for generative AI features on Google SearchAccessed Back to the text, first mention of source 5
  6. Google Search Central, Introduction to structured data markupAccessed Back to the text, first mention of source 6
  7. Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024, especially Appendix C.1Accessed Back to the text, first mention of source 7
  8. Bajemon and Rochet, Scoring Without the Engine, September 2026 preprint, section 8Accessed Back to the text, first mention of source 8
  9. Moore and Dunne, What Drives Citations in Production Large Language Models?, September 2026 preprintAccessed Back to the text, first mention of source 9
  10. Google Search Central, Spam policies for Google web searchAccessed Back to the text, first mention of source 10
  11. Google Search Console Help, Search generative AI controlAccessed Back to the text, first mention of source 11
  12. Google Search Console Help, Generative AI performance reportAccessed Back to the text, first mention of source 12
  13. Microsoft Bing, Introducing AI Performance in Bing Webmaster ToolsAccessed Back to the text, first mention of source 13