Why the Same Business Appears in ChatGPT for One User but Not Another

A business owner searches for their category in ChatGPT. Their business appears. Two weeks later, an audit of the same query says they don't. A competitor runs the same test themselves and gets a different list entirely. Nobody's lying. ChatGPT didn't break. There are two distinct mechanisms that explain this, and understanding them changes what an audit actually measures.

The reasoning mode you can't see from the outside

ChatGPT has two operating modes: standard reasoning and extended thinking. Free users run standard. Paying subscribers -- ChatGPT Plus and Pro -- can switch to extended thinking on any query. According to subscription data referenced in our July 10, 2026 methodology rec (`research-vault/methodology-recs/2026-07-10-chatgpt-reasoning-mode-audit-scope.md`, Scout session 73), somewhere between 20-30% of ChatGPT's monthly active users are paying subscribers. Every one of them can enable extended thinking for any search they run.

The two modes don't just produce slightly different answers. A June 30, 2026 Semrush study with researcher Kevin Indig, run across 100 prompts in 20 buyer journeys, found the following:

| Mode | Citation rate | Sources per response | Sub-queries per prompt | Web searches per 100 prompts | |---|---|---|---|---| | Standard | 50% | 2.6 | 5.5 | 245 | | Extended thinking | 68% | 4.5 | 24 | 1,130 |

Source overlap between the two modes: 25.6%. That means 74.4% of the domains cited in extended thinking don't appear in standard mode results for the same queries.

The July 10 rec documents our conclusion: these are not minor variations. Standard mode runs 5-6 sub-queries per prompt and returns 2-3 sources. Extended thinking runs 24 sub-queries and returns 4-5 sources from a pool that is nearly entirely different. A paying subscriber running a high-stakes query with extended thinking enabled is searching a version of the web that standard-mode users don't see.

When a business owner tests ChatGPT and finds their name, there is no way to tell from the outside whether they searched in standard or extended mode -- unless they check their interface settings. Most don't. Most assume the result they saw is the result everyone sees.

Why the same mode also varies between runs

Extended thinking versus standard mode is a structural divergence. The second mechanism operates within a single mode.

LLMs are non-deterministic. ChatGPT generates responses by sampling from probability distributions, and that sampling introduces randomness. The same query, same mode, same user, run a second time seconds later, can produce different citation outputs.

Our July 8, 2026 methodology rec (`research-vault/methodology-recs/2026-07-08-audit-measurement-non-determinism.md`, Scout session 71) synthesized an April 2026 academic paper that formalizes this problem for AI citation research: Schulte et al., "Don't Measure Once: Methodological Implications of LLM Non-Determinism for GEO Research" (arXiv:2604.07585).

The paper's core finding: a single-run measurement of citation presence produces an unreliable point estimate. Citation presence for a given business can flip between runs with meaningful frequency, specifically for businesses near the edge of the citation threshold -- the point where they appear in some configurations but not others. The paper argues that citation behavior should be modeled as a probability distribution across repeated runs, not as a binary present/absent state.

The July 8 rec maps this to Sourcepull's audit context: a business that appears in 1 out of 5 audit runs is not "cited" or "not cited." It has a 20% citation probability for that query in that mode. Reporting a single run as definitive obscures this.

What it looks like in practice

These two mechanisms combine to produce the scenario business owners find confusing.

A business with moderate citation signals -- enough to appear in some configurations, not in others -- might have a 30% citation probability in standard mode and a higher rate in extended thinking, simply because extended thinking's 24 sub-queries cast enough angles to catch the business in at least one of them. Standard mode's 5-6 sub-queries might not.

A paying subscriber who uses extended thinking for everything tests ChatGPT and finds their business. An audit run in standard mode (the correct methodology for measuring the majority user path) doesn't show them. Both are accurate. They're measuring different things.

The inverse also happens: a business with strong citation signals shows up in both modes, consistently across multiple runs. An independent contractor with thin directory presence and entity ambiguity shows up sporadically -- sometimes in extended thinking, almost never in standard.

This is why business owners occasionally report that they "saw themselves in ChatGPT" before a Sourcepull audit shows near-zero scores. The audit, per the July 10 rec, is deliberately run in standard mode because "auditing in high-reasoning mode would over-represent citations for the paid-subscriber minority and produce an inaccurate picture of the market's default user path." The subscriber who enabled extended thinking got a result from a 74%-different source pool.

What actually determines citation probability

Given that ChatGPT citation behavior is probabilistic and varies by mode, what moves the distribution in the right direction for both?

Citation signal strength is the underlying variable. Businesses that appear consistently across modes and repeated runs share specific characteristics: directory presence across sources ChatGPT retrieves from, clear entity signals (Wikidata entry, consistent NAP across platforms, category clarity), and third-party mentions in content that ended up in training data or retrieval pools.

Extended thinking, by running 24 sub-queries, effectively multiplies retrieval surface area. A business that appears in one of many query angles has a reasonable chance of showing up. But appearing in extended thinking because of a lucky sub-query match doesn't indicate strong citation presence -- it indicates a thin signal that got caught by one of many nets. Standard mode's narrower retrieval exposes the actual signal strength: does this business appear when the query is run with typical consumer-path parameters?

The Schulte et al. paper recommends measuring citation behavior across 3-5 repeated runs and reporting citation probability rather than a single point. Our July 8 rec identifies this as the directional change for how Sourcepull will frame re-audit results: trend direction across repeated measurements, not point comparison between single runs. A business that moves from a 20% citation rate to a 40% citation rate across multiple runs has made real progress, even if any single-run audit could report either outcome based on noise.

What this means before you test yourself

If you're planning to test your own ChatGPT citation presence, two things to keep in mind.

First, check which reasoning mode you're in. If you're a Plus or Pro subscriber and extended thinking is enabled, you're seeing results from a source pool that's 74% different from what standard-mode queries return. That's not your AI visibility score -- it's a higher-stakes retrieval that catches more signals than typical users encounter.

Second, run the same query multiple times. If your business appears in two out of five tries, you have a 40% citation probability in that mode. If it appears in one out of five, you have a 20% probability. The variance is real and meaningful, not a glitch.

A Signal Check at sourcepull.ca runs queries in standard mode across ChatGPT, Perplexity, Gemini, and Claude, and gives you per-platform citation rates. If the score is lower than what you see in your own manual testing, reasoning mode is the first thing to verify. For businesses building from low citations to consistent presence, that per-platform breakdown shows which source signals are missing -- because the citation probability in either mode is determined by the same underlying signals.