Two Schema Studies Got Opposite Results. Both Are Right.

Two of the most-cited schema studies in the AEO space reached opposite conclusions. OtterlyAI reported a 1,500%+ lift in Google AI Overviews appearances after schema implementation. Ahrefs ran the first controlled schema study and found no statistically significant change in AI citation rates on any platform -- and a significant negative result for AI Overviews.

Both studies are credible. Both measured the same variable. They reached opposite conclusions because they studied different populations.

Our August 2026 investigation resolved this conflict. The resolution changes how schema fits into an AEO fix plan -- and for some businesses, it means the schema work they're about to do is the wrong next move.

What the two studies actually measured

OtterlyAI's finding came from BrightonSEO 2025. They tracked pages before and after schema implementation and documented dramatic increases in AI Overviews appearances. The cases they cited involved businesses that were previously receiving essentially no AI citations. They were invisible before schema was added.

Ahrefs' controlled study -- which we documented in our research vault when it published (Scout session 26, 2026-05-23) -- took a different approach. They tracked 1,885 pages that already had 100 or more AI citations before schema was added. They matched these against 4,000 control pages with similar citation histories. The result: +2.4% on AI Mode (not significant), +2.2% on ChatGPT (not significant), -4.6% on AI Overviews (statistically significant, negative).

The key difference is the starting population. OtterlyAI tested pages that had no AI presence. Ahrefs tested pages that were already being cited at scale. These are structurally different experiments. The conflicting results are not contradictory -- they're measuring schema's effect at two distinct starting points, and the effect is different at each.

The threshold that separates the two findings

Our August 6, 2026 methodology update (Scout session 99, `methodology-recs/2026-08-06-schema-threshold-fix-plan-stratification.md`) formalized what this pattern implies:

Schema markup functions as a threshold mechanism, not a citation frequency booster.

For businesses with no AI citations, proper schema gives AI systems the entity infrastructure they need to classify, recognize, and begin citing the business. LocalBusiness schema with a complete address, consistent business name, correct category, and sameAs links to verified directories tells the model: here is an entity, here is what it does, here is where it operates, here are third parties that confirm this. That is information the model did not have. It changes what is possible for the model to cite.

For businesses already receiving citations -- already recognized and already in AI systems' consideration set -- adding schema does not give the model information it was missing. The entity is already classified. Adding FAQPage schema or expanding JSON-LD fields does not meaningfully change the probability of citation when a relevant query arrives.

There is a detail worth noting from the Ahrefs dataset itself: their broader 6M-URL correlation analysis found a 3x schema correlation before they filtered down to the already-cited study population. That pattern is consistent with the threshold hypothesis -- schema correlates with citation presence strongly across the full population, but the effect disappears when you isolate pages that were already cited. Two independent observations pointing to the same mechanism.

It is worth being direct about confidence level here. No controlled study has directly tested the threshold by running the same schema addition on uncited versus cited pages simultaneously. This is a hypothesis, well-supported by practitioner evidence and by Ahrefs' own broader data, but not confirmed by a head-to-head experiment. The August 6 rec explicitly rates this LOW-MEDIUM confidence. The stratified guidance that follows from it is a working framework, not a settled finding.

For businesses with no AI citation presence

If your business is receiving few or zero AI citations across all platforms -- you do not appear consistently in ChatGPT for your category, Perplexity is not citing your website, Claude is not mentioning your business in local recommendations -- schema is foundational work that belongs first.

The implementation that matters: LocalBusiness schema with your exact business name matching your GBP and Yelp listing, complete address fields, primary phone, business hours, and sameAs links to every verified directory listing you have claimed. The @id field should point to your canonical domain.

The August 6 methodology rec frames this as "entity establishment infrastructure -- required before other signals can work reliably." Directory presence, content freshness, and review signals all work better when the AI system has a confident, consistent entity record to anchor them to. Schema is not optional for businesses starting from zero AI presence.

For businesses already appearing in some AI searches

If your business is already visible in AI search -- you appear in some Perplexity queries, ChatGPT cites you for some category queries, Gemini mentions you -- schema is not where your next improvement will come from.

The August 6 methodology update was specific about priority order for already-cited businesses: directories and freshness are the primary levers. Schema belongs in a secondary pass, after the higher-return work is in place.

For Google AI Mode gaps, freshness is the primary signal. For Perplexity gaps, format matching and content specificity drive citation rate. For ChatGPT gaps, the platform-specific picture is more consequential.

Why ChatGPT schema becomes lower-return than anywhere else

Our August 2, 2026 methodology update (Scout sessions 95-96, `methodology-recs/2026-08-02-yelp-chatgpt-thumbtack-claude-partnerships.md`) documented a structural shift in how ChatGPT's local citation mechanism works.

ChatGPT local recommendations are now driven primarily by three direct data partnerships: Foursquare (December 2024), Thumbtack for home services (October 2025, confirmed from BusinessWire), and Yelp (July 2026, 330M reviews licensed to OpenAI). These are structured API feeds, not web retrieval. When ChatGPT answers "plumber near me" or "dentist in Hamilton," the local results come from those database integrations, not from crawling business websites or parsing schema markup.

Schema on your website affects what AI systems extract when they crawl your pages. Foursquare, Yelp, and Thumbtack feed ChatGPT through direct integrations. Your schema does not enter that pipeline.

The August 2 methodology update revised ChatGPT fix plan priority for local businesses:

- Priority 1 (data partnerships): Foursquare, Yelp, and Thumbtack for home services - Priority 2 (organic training-data presence): BBB, Angi - Schema: secondary pass, after data partnership listings are complete and accurate

A business spending time refining ChatGPT-specific schema while its Foursquare record is stale or its Yelp profile is incomplete is working in the wrong order. The data partnerships dominate ChatGPT's local citation pipeline in ways that schema cannot reach.

This applies to the August 6 Phase 2 guidance specifically: for already-cited businesses with ChatGPT gaps, schema is not just lower-priority -- it is "low-return post-GPT-5.5" per the methodology rec. The ChatGPT fix path bypasses schema entirely and goes straight to directory data quality.

The priority sequence

Schema establishes the entity foundation. This is not optional for businesses with no AI citation presence. The downstream fixes produce less reliable results without it.

After that, the paths diverge by platform and by where you are already visible. Directory presence and freshness drive citation frequency for already-cited businesses. Schema refinement does not. For ChatGPT specifically, data partnership listings should be complete and verified before schema optimization is prioritized.

The threshold question for any business is simple: are you invisible to AI, or are you appearing inconsistently? Invisible means schema first. Inconsistent means directories and freshness first, with schema as a cleanup pass.

If you are not sure which category describes your business, Signal Check at sourcepull.ca runs a live audit across ChatGPT, Perplexity, Claude, and Gemini. The score shows where you currently stand -- and the fix plan that follows is stratified by what that score reveals, not by a one-size-fits-all schema checklist.