Schema Markup for AI Citations: The Platform Split That Changes Everything

In June 2026, we published our analysis of the Ahrefs and Fischman schema studies and explained why they appeared to contradict each other. The short version: Ahrefs tested businesses already receiving citations (schema addition didn't move them higher), while Fischman tested lower-authority sites across the full citation range (attribute-rich schema correlated with a 22-point citation rate increase). The studies measure different populations. Both are right.

That reconciliation holds. But since June, our vault research has accumulated enough platform-specific data to add a third dimension to the schema question -- one that changes the operational answer depending on which AI platform you're trying to fix.

The Phase 1/Phase 2 Foundation (Still Valid)

Our methodology recommendation filed after Scout session 99 (2026-08-06) formalizes the two-phase framework in fix plan language:

Phase 1 clients -- businesses with few or zero AI citations across all platforms -- should treat schema as entity establishment infrastructure. The goal is to help AI systems classify and discover the business for the first time. LocalBusiness schema with a complete address, a correct `@id`, and a `sameAs` array pointing to your claimed directory listings gives retrieval systems something anchored and consistent to work from. Without it, a business can be invisible not because it's unknown but because nothing names it unambiguously.

Phase 2 clients -- businesses receiving some AI citations but underperforming on specific platforms -- shouldn't lead with schema. The fix plan rec is explicit on this: "the primary levers for already-cited clients are freshness, format matching (Perplexity), directory presence (ChatGPT), and GBP completeness (AI Mode). Schema work belongs in a secondary pass after the directory and content actions are in place."

This is the same conclusion the June analysis reached via the academic studies. The August update makes it operational: schema has a specific place in the fix sequence, and it isn't first for Phase 2 clients on any platform.

For ChatGPT, the Analysis Goes Further

The Phase 1/Phase 2 split applies cleanly to Perplexity and Google AI Mode. For ChatGPT, something different happened over the past 12 months that makes schema a lower-return action regardless of which phase a client is in.

Our methodology rec filed after Scout sessions 95-110 (2026-08-02 through 2026-08-17) documents three confirmed data partnerships that now feed ChatGPT local results directly:

**Foursquare** (December 2024): ChatGPT's local POI database. A Yext study of 6.8 million ChatGPT citations found 60-70% of local business results trace back to Foursquare. This is a background data feed -- the mechanism that determines which businesses appear in ChatGPT local responses before any schema signal is evaluated.

**Thumbtack** (October 2025, home services): Thumbtack marketplace professionals are served directly in ChatGPT via API when users ask home-related questions. A business absent from Thumbtack is invisible to this layer regardless of how complete its schema is.

**Yelp** (July 2026): Yelp licensed 330 million reviews and 8 million business listings to OpenAI via structured API. Yelp profile completeness is now a direct ChatGPT signal -- not an organic citation opportunity, but a data feed that populates ChatGPT's local response layer.

The practical consequence, from the session 99 rec: "For ChatGPT gaps: schema is a low-return action post-GPT-5.5. Third-party directory presence and Foursquare/Yelp data quality are more effective levers." A business can have correct, attribute-rich LocalBusiness schema and still be structurally absent from ChatGPT local recommendations if its Foursquare listing is unclaimed, its Yelp profile is thin, or (for home services) it has no Thumbtack listing.

This isn't a claim that schema harms ChatGPT performance. A clean schema implementation remains part of the infrastructure baseline. The claim is about priority: for ChatGPT, fixing directories and data partnership presence should come before any schema work, and for a Phase 2 client in particular, schema is likely not the explanation for ChatGPT gaps at all.

Perplexity and Google AI Mode: Phase 1 Schema Still Matters

Both Perplexity and Google AI Mode perform live web retrieval when generating answers. They crawl pages, evaluate content structure, and make real-time citation decisions -- which means on-site signals including schema markup remain in the signal mix.

For Phase 1 clients targeting these platforms, the August 6 rec is clear: schema is high priority within Phase 1 entity establishment actions. The framing is "entity establishment infrastructure -- required before other signals can work reliably." For Phase 2 clients on these platforms, the priority drops but doesn't disappear: ensure schema is correct, then move to platform-specific levers (freshness and format for Perplexity, GBP completeness for AI Mode).

For home services clients specifically, the August session 110 finding adds a Gemini dimension: Angi and Thumbtack joined Gemini as Connected Apps on August 12, 2026. For Gemini, these integrations now handle home project routing at the action layer -- which means the same directory-first logic that applies to ChatGPT now applies to Gemini for this category. Schema remains infrastructure, but the Connected App routing layer operates independently of schema.

Citation Stability Is a Separate Problem From Entry

Our methodology rec filed after Scout session 92 (2026-07-30) introduced the SISTRIX citation drift data. SISTRIX tracked 82,619 prompts and 1.5 million citation snapshots across 17 weeks. The headline numbers: Google AI Mode rotates 56% of its cited sources each week; ChatGPT rotates 74%.

Schema can help a Phase 1 business enter the citation pool. It does not keep it there. The stability data shows that brand domains -- businesses with consistent directory presence, fresh content, and accurate entity signals across multiple sources -- hold in the 43% of sources that remain stable across all 17 weeks. The other 57% rotates in and out weekly.

The operational implication: schema is entry infrastructure, not retention infrastructure. Sustained citation presence requires the signals that AI retrieval systems re-evaluate continuously -- directory accuracy, review freshness, NAP consistency, and (on Google AI Mode) GBP completeness. A one-time schema implementation gets you into the candidate pool. The directory and content work keeps you in it.

The Fix Sequence

Phase 1, any platform: schema as entity establishment (LocalBusiness with `@id`, address, and `sameAs` pointing to claimed directory listings) is a prerequisite alongside Foursquare and Yelp claims. For ChatGPT, the directory claims are more urgent than the schema. For Perplexity and AI Mode, both are roughly equal priority.

Phase 2, ChatGPT gaps: schema is not the likely explanation. Check Foursquare claim status, Yelp profile completeness and review volume, and Thumbtack listing (home services). For ChatGPT at this stage, directory infrastructure is the primary lever.

Phase 2, Perplexity gaps: freshness and content format are the primary levers. Schema is worth verifying is correct but unlikely to be the gap driver.

Phase 2, Google AI Mode gaps: GBP completeness and Yelp presence are the primary levers. Schema should be correct and complete, but it doesn't explain most AI Mode citation gaps at Phase 2.

If you want to see which platforms are citing you and which directories are missing from your data footprint, Signal Check at sourcepull.ca runs a live diagnostic across ChatGPT, Perplexity, Gemini, and Google AI Mode and shows exactly where the gaps are -- including whether schema issues or directory gaps are the more likely explanation for what you're seeing.