Short answer: ChatGPT and Perplexity overlapped only partially in the measured prompt set. In Nalmert Research 001, the same frozen prompt set was measured in ChatGPT and Perplexity across four markets. ChatGPT completed 36 observations and Perplexity completed 35; one UK Perplexity observation remained a documented technical failure.
This comparison describes provider mentions, explicit recommendations and visible citations in those measured responses. It does not rank provider quality, test campaign performance or determine which AI system is “better.” Results are time-, session-, interface- and market-dependent.
What was compared
Research 001 used 36 frozen prompts covering nine commercial-intent categories across International English, the United States, the United Kingdom and Russia. ChatGPT and Perplexity received equivalent prompts for each planned cell. Alice was measured separately for nine Russian observations and is not included in the comparison below.
The unit behind the mention and recommendation counts is a validated observation–provider row. A provider can therefore appear once in several different observations. Citation counts use a different unit: one observation plus one visible citation URL. Neither unit is a quality score.
ChatGPT vs Perplexity: comparison at a glance
| Metric | ChatGPT | Perplexity | How to interpret |
|---|---|---|---|
| Completed observations | 36 | 35 | One planned UK Perplexity observation failed twice and was not measured. |
| Validated provider mention rows | 27 | 259 | An observation–provider pair in which the provider appeared. Repeated appearances across observations remain separate rows. |
| Explicit recommendation rows | 27 | 73 | A subset of mention rows where the response used a clear recommendation or shortlist role. |
| Unique normalized provider entities mentioned | 23 | 202 | Distinct normalized provider entities observed at least once in that engine’s measured responses. |
| Visible citation rows | 337 | 176 | Visible source URLs recorded from the interface. A citation is not automatically evidence that the cited company was recommended. |
The large differences in row counts describe the outputs captured in this bounded measurement wave. They do not show that one engine is more accurate, more trustworthy or generally more likely to recommend a provider.
Mention, recommendation, citation and engine coverage
Mention means a normalized provider entity appeared in the answer. The provider may have been listed, described, compared or referenced without being endorsed.
Explicit Recommendation means the answer placed the provider in a clear recommendation, shortlist or equivalent prescriptive role. Every explicit recommendation is also a mention, but not every mention is a recommendation.
Citation means a visible source URL was exposed with the response. The cited source may support a provider statement, a general claim or background context. A citation does not by itself mean that the cited company was the provider being recommended. The association-level problem is examined separately in Research 003.
Engine Coverage records whether a provider appeared at least once in each measured engine. Coverage of two means presence in both ChatGPT and Perplexity; it does not mean equal frequency, equal recommendation strength or a stable ranking position.
The core rule remains: Mention ≠ Recommendation ≠ Citation.
Where ChatGPT and Perplexity differed
The measured responses contained different volumes of provider entities
The validated ChatGPT records contained 27 provider mention rows across 36 completed observations. The Perplexity records contained 259 mention rows across 35 completed observations. In this wave, ChatGPT’s captured responses produced a smaller provider set, while Perplexity’s captured responses contained many more provider entities.
That is an observation about these interfaces and sessions, not a universal property of either system. The study did not experimentally isolate answer length, retrieval mode or interface design as causes.
Recommendation counts did not move in parallel with mention counts
All 27 validated ChatGPT provider rows met the study’s explicit-recommendation rule. Perplexity produced 73 explicit recommendation rows from 259 mention rows. These counts describe how providers were framed in the measured answers; they do not measure whether the recommendations were useful or commercially appropriate.
Citation volume followed a different pattern again
The evidence archive contains 337 visible ChatGPT citation rows and 176 visible Perplexity citation rows. Citation volume must remain separate from provider visibility: one source can support general context, several providers or a statement unrelated to a recommendation. “Third-party” also does not automatically mean independent editorial endorsement.
What cross-engine overlap means
Sixteen normalized provider entities appeared in both the ChatGPT and Perplexity datasets. Seven appeared only in ChatGPT and 186 appeared only in Perplexity. The union across the two engines contained 209 normalized provider entities.
Expressed as a Jaccard overlap, the result is 16 ÷ 209 = 7.7%. The denominator is the union of provider entities observed in either engine, not the number of prompts or responses. The metric describes set overlap only. It is not a quality score, recommendation rate or estimate of future visibility.
Among the more frequently surfaced providers with engine coverage of two were iPullRank, Head Promo, Found, Go Fish Digital, Digital Geeks, Directive Consulting, Rank4AI and Seer Interactive. Other providers appeared in only one measured engine. Because the long tail includes many entities observed once, the overlap percentage is sensitive to list length and should be interpreted together with the underlying counts.
Why equivalent prompts can produce different answers
Several methodological factors can contribute to cross-engine variation:
- the engine and interface mode used for the observation;
- which sources were available or retrieved at that moment;
- wording and commercial intent of the prompt;
- market and language context;
- changes over time, sessions and product updates;
- differences in how an answer is composed and how citations are exposed.
Research 001 observed different provider sets. It did not isolate these factors experimentally, so none should be treated as the proven cause of a specific provider’s presence or absence.
If your brand appears in one engine but not another
Treat the difference as a diagnostic signal, not as proof of an optimization effect. A practical review can check:
- Prompt coverage. Was the brand evaluated across enough relevant questions, markets and customer contexts?
- Entity clarity. Is it unambiguous who the company is, what it provides and which markets it serves?
- Source availability. Are accurate provider-owned and third-party sources available for the claims users are asking about?
- Content relevance. Do relevant pages directly answer the commercial questions represented by the frozen prompt set?
- Repeated measurement. Does the pattern persist when the same protocol is rerun at a later checkpoint?
These checks may help explain where evidence is missing. They do not guarantee a mention or recommendation. Nalmert’s AI Visibility framework and measurement methodology describe how to preserve prompts, zero results, markets and observation evidence.
Limitations
- This is a secondary comparison of Research 001, not a new measurement wave.
- The samples are small and interface-dependent: 36 completed ChatGPT observations and 35 completed Perplexity observations.
- One planned UK Perplexity observation was not measured after two technical failures.
- Provider rows within one response are not independent statistical trials.
- The long tail of normalized entities affects unique-provider and overlap counts.
- Visible citations reflect what the interfaces exposed, not every source that may have influenced an answer.
- The study did not test provider quality, campaign outcomes or causal optimization effects.
- Results should not be generalized to all prompts, dates, users, markets or interface modes.
Frequently asked questions
Do ChatGPT and Perplexity recommend the same GEO providers?
They overlapped only partially in the measured prompt set. Sixteen normalized provider entities appeared in both engine datasets, while other providers appeared in only one. The overlap describes this bounded wave and should not be treated as a permanent provider ranking.
Does being mentioned mean a provider was recommended?
No. A mention records provider presence. An explicit recommendation requires clear prescriptive or shortlist language. The two signals are stored separately.
Why can a provider appear in one AI system but not another?
Answers can vary with engine, interface, retrieved sources, prompt wording, market, language, time and session. Research 001 observed the differences but did not isolate which factor caused any individual result.
Does a citation mean the cited company was recommended?
No. A citation is a visible source URL. It can support a statement about another provider or provide general context. Citation ownership and the relationship to a recommended provider require association-level review.
Which engine is better for GEO provider recommendations?
This study was not designed to select a better model or interface. It measured provider presence, explicit recommendations and visible citations under a frozen protocol. Usefulness, factual accuracy and provider quality require different evaluation designs.
Should AI visibility be measured in more than one engine?
Yes when the research question concerns cross-engine visibility. The incomplete overlap in this wave shows why a result from one engine should not automatically be generalized to another. Any comparison should preserve equivalent prompts, markets, dates and evidence.
Methodology and primary source
Review the complete Research 001 dataset, findings and limitations before interpreting this comparison. The Research page is the primary source; this Insight explains one cross-engine slice of the same evidence.