Two Real Signals, One Myth: What Actually Moves ChatGPT Citations

A claim circulating in AEO vendor content goes something like this: once an answer pattern for a commercial query gets cached, the platform has a direct cost incentive to reuse it. Brands that enter those early cached answers gain a moat backed by compute economics. Move fast, get in now, structural advantage secured.

We investigated this claim in detail. It is wrong -- not in a subtle way. Understanding why matters because it changes where you should be spending your fix budget.

The Myth: Prompt Caching as a Citation Moat

Prompt caching in the OpenAI API is a real feature. Introduced in October 2024 and updated in May 2026, it gives developers a 90% input token discount when the beginning of their API request matches a previously processed prompt. It is a cost-reduction tool for companies building applications on top of GPT -- it reduces what developers pay OpenAI to process repetitive prefixes like system prompts and conversation history.

Scout's 2026-07-11 edge-case analysis traced the technical mechanism in full detail after a vendor published the cached-answers claim with specific numbers attached. The conclusion was unambiguous: prompt caching discounts the cost of reading back a previously computed representation of a developer's own prompt prefix. It applies to input tokens. It has no mechanism that would cause ChatGPT to prefer citing any specific business or source across multiple users' queries.

ChatGPT's consumer interface generates a fresh response to each user query. Whether it cites your business when someone asks "best plumber in Burlington" is governed by the model's training data and its real-time web retrieval -- not by a cost-optimization layer built for API developers. There is no shared citation cache across users. OpenAI does not benefit financially from reusing your brand in answers to different users.

What vendors may be imprecisely gesturing at is real: brands that appear more frequently in training data and that consistently rank in web search results have more durable citation presence. That is not caching. That is training data prevalence and retrieval stability -- two different mechanisms, each with its own fix path.

Real Signal 1: Structural Data Feed Integrations

Our 2026-08-02 methodology rec (Scout session 95) documented three partnerships that now structurally control which home services businesses appear in ChatGPT and Claude local responses -- partnerships confirmed from primary press releases, not from vendor claims.

**Yelp plus OpenAI (July 23, 2026):** Yelp licensed 330 million reviews and 8 million business listings to OpenAI. This data now feeds ChatGPT local responses via a direct API -- not organic training data, not web crawl. A dedicated data feed. The implication is stark: a business absent from Yelp is structurally absent from ChatGPT's local data layer regardless of website quality, schema markup, or content depth.

**Thumbtack plus OpenAI (October 2025) and Thumbtack plus Claude (April 2026):** Thumbtack's marketplace was integrated directly into ChatGPT in October 2025, then into Claude in April 2026. For home services queries, a business not listed on Thumbtack is invisible in both platforms' response layers. This is not a citation probability gap -- it is a structural exclusion that website-level optimizations cannot compensate for.

The updated priority order for ChatGPT (home services) that came out of this research:

- Priority 1, data feed partnerships: Foursquare, Thumbtack, Yelp - Priority 2, organic citation sources: BBB, Angi

If your fix plan does not distinguish between data feed partnerships and organic citation sources, it is treating structurally different signals as equivalent. They are not. Fixing your schema while your Yelp profile is incomplete or unclaimed leaves the structural gate closed.

Real Signal 2: Schema as Entity Establishment, Not Frequency Booster

The second misalignment we see in AEO advice is treating schema markup as a universal lever for all businesses regardless of where they are in the visibility spectrum. Scout's 2026-08-06 methodology rec (session 99) investigated why two major studies reached opposite conclusions on schema's impact -- and resolved the conflict at a functional level.

The OtterlyAI BrightonSEO study found a +1,500% AI Overviews increase after schema implementation. The Ahrefs study found a null result. Both are correct because they tested structurally different populations.

The Ahrefs data is the tell: a 3x schema correlation appeared in their full six-million-URL population before they narrowed to pages already receiving AI citations. The correlation collapsed once they looked only at already-cited pages. The threshold hypothesis that emerged: schema helps uncited pages enter the citation pool; it does not increase citation frequency for pages already in it.

For businesses with few or zero AI citations across all platforms, schema markup is entity establishment infrastructure. LocalBusiness schema with correct address fields, a proper @id, and sameAs links to claimed directory listings is foundational -- not because schema directly lifts citation frequency, but because its absence prevents the AI systems from classifying and discovering the business in the first place.

For businesses already receiving AI citations but with gaps on specific platforms, schema is an amplifier, not a primary lever. The primary fix levers at that stage are directory presence, content freshness, and format matching. Running another schema optimization pass when directories are the actual gap is a confident action in the wrong direction.

Where to Put the Fix Budget

These three findings together resolve a common prioritization failure in AEO planning.

Time spent trying to enter cached answer pools is wasted -- there is no citation-level cache accessible to businesses or agencies. For home services businesses, directory presence on Foursquare, Yelp, and Thumbtack is not a supporting action. It is the structural gate. A business absent from these platforms is excluded from the data feeds that actually power platform responses, regardless of what its website says.

Schema work is most valuable for businesses that are not yet in the AI citation pool. For businesses with existing citations and specific platform gaps, the same effort applied to directory completeness and content structure is more likely to produce measurable change.

If you are not sure where your business sits on that spectrum -- uncited, partially cited, or cited with specific platform gaps -- that is exactly what a Signal Check surfaces. Free, three minutes, gives you your current AI citation status across platforms so you know which phase of fixes applies to you.