ELNIQ GEOQueseo
  1. Home
  2. GEO study

GEO study · Valencia · 2026

One question, many lists

How recommendations of SEO consultants in Valencia change when you ask ChatGPT the same question again and again. We asked the question 200 times, half in free ChatGPT and the other half through the OpenAI API, and counted which names came up, in what position and with what sources.

ZIP · 9 MB · study in PDF in five languages, raw data and preliminary analyses · free and no sign-up

  • 200attempts
  • 198valid lists
  • 26candidates
  • 0identical lists across environments

Nina S. Dzhezher (ELNIQ) and Eric Jorge Seguí Parejo (Queseo)Data from 8 and 9 November 2026

0%Download the ZIP
Chapter 一

The question

01 / 15 · Summary

What this test shows

If you ask ChatGPT for the best SEO consultants in your city, you see a list of five names. That list looks like an answer, but it is one among many possible ones. In this study we repeated the same question 200 times and the lists almost never matched. The names changed, the order changed and the distribution changed between ChatGPT in the browser and the OpenAI API.

  • 94different lists in the 98 browser responsesA list counts as different if any name or the order changes. In the API there were 95 in 100 responses.
  • 1,77names shared on average between a browser list and an API listOut of a maximum of five.
  • 26different candidates in the 198 lists17 appear in both environments.

These are the five main findings:

  1. Finding 1The list almost never repeats. In 98 browser responses there were 94 different lists, and in 100 from the API, 95. A single changed name or a change in order is enough for the list to be a different one.
  2. Finding 2The two environments show different options. When they were compared, there was no list identical in names and order. Enrique Tejedo appeared in 77 of the 98 browser lists and in one of the 100 API lists. Queseo went the opposite way, with 37 of 98 and 87 of 100.
  3. Finding 3Appearing and coming first are different measures. Queseo was the most frequent brand in the API, but Jesús Porres came first in more lists.
  4. Finding 4Sources and listings go hand in hand with other distributions. There are clear associations between what ChatGPT consults and whom it recommends, although this test does not isolate their causes.
  5. Finding 5Content is also part of visibility. It matters which specialism, which links and which reviews ChatGPT associates with each name.
The video is on the ELNIQ GEO YouTube channel. ELNIQ also publishes its version of the study on its blog.

02 / 15 · Method

How it was done: the same question in two environments

We always asked the same question, in the same words:

«Recomiéndame 5 consultores SEO en Valencia, España. Explica brevemente por qué recomiendas a cada uno.»

We sent it in two ways. In the browser, from the ChatGPT website, as anyone would. In the API, which is the route a program uses to send a question to OpenAI and receive the answer.

ConditionBrowserAPI
AccessFree ChatGPT, not logged in. New chat for each attempt.Requests from a program, with web search enabled.
Attempts and valid lists100 attempts; 98 lists.100 attempts; 100 lists.
ModelNot identified.gpt-5.6-luna, requested and returned in the log.
Time (mainland Spain)8 November, 21:34, to 9 November, 00:35.8 November, from 21:32:59 to 22:03:42.
Information keptText, site and page records, listings and response elements.Text, links, usage and number of search calls.

There were 2 failed attempts in the browser, numbers 28 and 44. They stalled on the search message and returned no list, so they are not included in the count.

What counts as a candidate

We counted the five numbered names in each main list and grouped together the variants that identify the same person or brand. For example, “Eric Jorge Seguí Parejo” and “Queseo” count as the same candidate. This gives 490 positions in the browser and 500 in the API. The candidates include independent professionals and teams.

Chapter 二

Every response is a new roll of the dice

03 / 15 · Variation

When the question is repeated, the names and the order change

There are two ways of comparing lists. A combination contains the same five names even if their order changes. An ordered list also requires each name to hold the same position. Both measures tell the same story, that the response varies a great deal from one attempt to the next.

What was countedBrowser (98 lists)API (100 lists)
Different candidates23
Combinations of five, ignoring order67
Different lists taking order into account95
Responses whose ordered list does not appear again90

Chart 1 · Repetition

How many names two responses share

Average number of names in common between two lists, out of a maximum of five

012345Two from the browser2,30Two from the API2,62One from each environment1,77
All possible pairs were compared: 4,753 within the browser, 4,950 within the API and 9,800 when crossing both environments. When they were crossed, there was no list identical in names and order.

Looked at pair by pair, the result is even clearer. Of the 4,753 pairs of browser responses, 47 have the same five names and only 5 also keep the order, roughly 0.1%. In the API the figures are 59 and 5. When both environments are crossed, 10 pairs share all five names and none shares the order.

MeasureBrowserAPICross
Pairs of responses compared4.9509.800
Average names shared2,621,77
Matches in name and position, on average0,810,32
Pairs with the same five names5910
Pairs with an identical ordered list50
Pairs with the same first name1.664729
Average Jaccard index0,3780,234
The Jaccard index divides the shared names by all the distinct names in the two lists. It ranges from 0 to 1.

Are two consecutive responses more alike than any two responses? Barely. Two consecutive browser lists share 2.29 names on average, compared with 2.30 across all their pairs. In the API the figures are 2.56 and 2.62.

Chapter 三

The twist: each environment shares things out differently

04 / 15 · Distribution

Whom each environment gives visibility to

Presence is the percentage of lists in which a candidate appears, in any of the five positions. We compare percentages because there are 98 valid lists in the browser and 100 in the API.

Chart 2 · The twist

Presence in the browser versus the API

Percentage of lists in which each candidate appears. The figure on the right is the difference of the API compared with the browser, in points.

BrowserAPI

0 %25 %50 %75 %100 %Enrique Tejedo-77,6David Cerdá (Seomar)-28,4Alex Figueiras+6,9Jesús Porres+33,0Queseo+49,2Rocío Santamaría+10,3
The five most frequent candidates in each environment, six when combined.

Queseo appeared in 37 of 98 browser lists (37.8%) and in 87 of 100 API lists (87.0%), a difference of 49.2 points. Enrique Tejedo went the other way, from 77 of 98 lists in the browser to one of 100 in the API. Alex Figueiras was the most stable of the six, with 54 of 98 and 62 of 100.

Chart 3 · Distribution

The 14 candidates with the most presence

Percentage of valid lists in which each one appears

Browser (98 lists)API (100 lists)

0 %25 %50 %75 %100 %Queseo37,8 %87,0 %Jesús Porres49,0 %82,0 %Alex Figueiras55,1 %62,0 %Enrique Tejedo78,6 %1,0 %David Cerdá (Seomar)72,4 %44,0 %Rocío Santamaría35,7 %46,0 %Álvaro Pichó (SEO Levante)36,7 %12,0 %Rafa Sospedra14,3 %42,0 %Alfredo Río (SEO Local Valencia)29,6 %1,0 %Andrés Castillo (GEO Valencia)16,3 %24,0 %Miguel Valero1,0 %26,0 %Paul Cris (Paul Marginean)22,4 %1,0 %Jamil Mansilla8,2 %23,0 %Paco Rubén16,3 %17,0 %
The 26 candidates, with their presence and the number of times they came first, are in the appendix table further down this page. The full distribution from first to fifth position is in appendix A of the PDF.

05 / 15 · Position

Appearing often is not the same as appearing at the top

Being in a list and holding its top positions are two different aspects of visibility. Position refers to the order of the text ChatGPT wrote, from first to fifth.

CandidateFirst in browserFirst in APITop 3 browserTop 3 API
Enrique Tejedo11
David Cerdá (Seomar)629
Alex Figueiras227
Jesús Porres5074
Queseo2877
Rocío Santamaría027
Number of responses, out of 98 in the browser and 100 in the API. “Top 3” includes first, second and third positions.

Queseo was the most repeated brand in the API and was among the top three in 77 responses, but Jesús Porres headed more lists, 50 versus 28. In the browser, Queseo appeared 37 times, ten of them among the top three and none in first position.

Alex Figueiras is the other example showing that frequency does not explain everything.

He appeared in 54 browser lists and in 62 API lists, but opened 25 of the former and only two of the latter. First position was shared among seven candidates in the browser and among nine in the API.

Chapter 四

More names than fit in one list

06 / 15 · Accumulation

Five options do not exhaust the set

The 198 lists contain 26 different candidates. Of these, 17 appear in both environments, three only in the browser and six only in the API. The more responses you review, the more names you discover that were missing from the first one.

Chart 4 · Accumulation

How many different names you see as you review more lists

Expected average number of different candidates by number of lists reviewed

BrowserAPI

05101520131020Lists reviewed5,09,414,216,45,08,913,215,4
Calculated on the saved lists, choosing them at random without repeating a response. It does not predict how many new names ChatGPT will offer in the future.

Even so, presence is concentrated. The nine most frequent candidates account for 83.5% of the 490 browser positions, and in the API the eight most frequent account for 82.6% of the 500. The combination made up of the five most frequent in each environment appears only three times in the browser and five in the API.

07 / 15 · Co-occurrences

There are patterns within the variation

The lists change, but some pairs of candidates tend to appear together more often than others. These examples come from the 98 browser lists:

Two candidatesLists in which they appear together
Queseo and Rocío Santamaría
Alex Figueiras and Álvaro Pichó
Paco Rubén and Alfredo Río
Paco Rubén and Alex Figueiras

Queseo appears in 37 browser lists and Rocío Santamaría in 35, and they coincide in 24. The API shows a different distribution, where Queseo and Jesús Porres coincide in 72 of its 100 lists.

The recorded sources also vary. Among the 97 browser responses that keep the site record there are 80 different sets of domains. If we rank the pairs of responses by how similar their sources are, the most similar third shares about 2.8 candidates and the least similar third, about 1.8.

Chapter 五

What ChatGPT reads before recommending

08 / 15 · Format

What the responses look like

The format is stable even though the names change. The browser usually presents criteria, a description of each candidate, a comparison and suggestions depending on the project. The API gives a shorter list, with links and guidance for choosing.

  • 582median words in the browserThe average is around 575.
  • 405median words in the APIThe average is around 400.
  • 98 de 98browser responses with a comparison tableIn the API, 0 of 100.
Element observedBrowserAPI
Comparison table98 of 980 of 100
Image elements in the raw text72 of 97No equivalent visual record
Visible map32 of 98No equivalent visual record
Literal phrase “no es un ranking” (it is not a ranking)23 of 9886 of 100

Why it recommends each one

Each list includes five explanations. We analysed their text looking for topics. They are mentions, not checks that a service or a credential is genuine.

Chart 5 · Topics

What ChatGPT talks about when it recommends

Percentage of descriptions that mention each topic. Bases of 490 descriptions in the browser and 500 in the API.

BrowserAPI

0 %25 %50 %75 %100 %Local SEO53 %47 %Length in years or “since”46 %43 %Technical SEO45 %45 %AI, GEO and generative search42 %51 %Online shops37 %45 %Number of reviews26 %7 %Case studies, testimonials or portfolio3 %8 %
A description can touch on several topics.

Visibility does not end with a name appearing. It also matters which need ChatGPT links it to, whether a migration, a local business, an online shop or a multilingual project. And each brand keeps its own associations even as it moves in and out of the lists. ChatGPT linked Queseo, for example, with migrations in all 37 browser descriptions and all 87 API descriptions.

09 / 15 · Sources

Where the information comes from

It is worth telling four things apart: a site appearing in the search record, being cited alongside a claim, being linked in the response and its author being recommended. The IP10 case shows this well.

  • 96 de 97browser site records include ip10.esIts Valencia comparison appears in the 97 page records.
  • 68 de 98visible texts with the source label “IP10”The direct link appears in 3 of 97 raw texts.
  • 0 de 198lists in which IP10 is a candidatePublishing a comparison and being recommended are different things.

A website can appear without being recommended

Each browser site record contains between seven and thirteen domains, with an average of 10.06. The five most frequent are ip10.es (96), valencianamente.es (90), albertofuster.com (79), sergidoseo.com (72) and jesusporres.com (68). For some candidates, having their website in the record is closely linked to appearing as recommended.

CandidateAppears if their site is recordedAppears if it is not
Álvaro Pichó
Alex Figueiras
Jesús Porres
Paul Cris
Andrés Castillo
The 97 browser responses, split into two groups for each candidate.

The relationship does not apply equally to everyone. Queseo and Rocío Santamaría appear frequently even though their domain barely features in that record, and Alberto Fuster is the opposite case: his domain is recorded 79 times and he appears as a candidate in only two. There are also professionals with several websites. Enrique Tejedo, for example, is linked to kenjimarketing.es in 44 responses and to seotejedo.es in two.

10 / 15 · Links

In the API, each candidate came with a link, five per response. Of those 500 links, 478 (95.6%) point to a domain identified with the recommended candidate. The remaining 22 go to third parties, ten of them to IP10, spread across four responses.

Linked domainBrowser (97 raw texts)API (100 responses)
queseo.es82
jesusporres.com82
alexfigueirasseo.com62
rafasospedra.com40
seomar.es40
rociosantamaria.com39
jamilmansilla.com22
Responses with at least one direct link to the domain. These are not visits or searches. The API adds utm_source=openai to 408 of its 500 links.

In the browser there are also business listings and maps. In 40 of the 98 responses some listing was saved, with data such as the rating, the reviews or the website, and 32 showed a visible map. The two signals coincide in 31 responses, so it is not correct to speak of 40 maps.

CandidateWith listing (40 responses)Without listing (58 responses)
Enrique Tejedo
Alfredo Río
Álvaro Pichó
Jesús Porres

The distribution changes between those groups, so it is advisable to record listings too when measuring visibility. The association does not prove that showing a listing brings a candidate in or takes them out, because both things can be part of the same selection.

11 / 15 · Discrepancies

The records make it possible to detect associations worth reviewing. For Jamil Mansilla, the recorded listing associates him with cristaleriasmansilla.com in five responses, while the link in the main text leads to jamilmansilla.com. The listings in responses 75 and 80 associate Rocío Santamaría with dramelguizo.com and 229 reviews, while the text of response 80 links to rociosantamaria.com and mentions 22 reviews.

With the files available, it is not possible to tell in every case whether the mix-up arose in the interface, in the selection of the listing or in its extraction. Nor do they prove a commercial connection between the businesses.

Review figures deserve the same caution.

In the first API response, 22 reviews are attributed to Rocío Santamaría and 14 to Jamil Mansilla. Both figures match the IP10 comparison, which is consistent with that origin but does not prove it. ChatGPT sometimes attributes the data (“according to their website”, “states”) and at other times asserts it directly. In API response 48, for example, it attributes more than 16 years of experience to Queseo. Attributing a piece of data shows who is claiming it, but it is not the same as verifying it.

Chapter 六

How to measure it without fooling yourself

12 / 15 · Sample

How many responses you need to measure

To measure a brand’s presence, we count how many lists it appears in and divide by the valid lists for that environment. That percentage describes the saved responses exactly. If we want to estimate how often the brand would appear if the test were repeated under the same conditions, there is uncertainty, because another batch could give a different percentage.

Chart 6 · Margin

95% margin by number of responses

Hypothetical example with a presence of 50%, the case with the greatest variation

0 %25 %50 %75 %100 %10 resp.±31.0 points30 resp.±17.9 points50 resp.±13.9 points100 resp.±9.8 points385 resp.±5.0 points
Calculation: 1.96 × square root of (0.25 / n), expressed in points. It assumes independence between responses and a stable frequency.

With 100 responses, the reference margin is ±9.8 points. Applied to a presence of 50%, the range runs from roughly 40.2% to 59.8%. With 385 responses it narrows to about ±5 points, from 45% to 55%. With a small batch, a single response moves the percentage a lot.

Chart 7 · Interval

Queseo’s presence with its 95% interval

Observed point and indicative range calculated with the Wilson method

0 %25 %50 %75 %100 %Browser37,8 %28,8 %47,6 %API87,0 %79,0 %92,2 %
37 of 98 lists in the browser and 87 of 100 in the API. These ranges do not predict the next response or the next week.

The presence of some candidates did change within the batch itself. Rocío Santamaría appears in 14 of the first 49 browser lists (28.6%) and in 21 of the last 49 (42.9%), and Andrés Castillo goes in the API from 15 of 50 lists (30.0%) to 9 of 50 (18.0%). That split does not make it possible to separate the variation between responses from a possible change in the system.

13 / 15 · What to do

How to run GEO tracking you can interpret

If you want to know whether ChatGPT recommends your business, a single question is not enough. These are the six steps we propose based on the test:

  1. Define which use you want to observe. Keep the question word for word and specify the environment, for example free ChatGPT in the browser without logging in, or a specific API configuration.
  2. Repeat, and record the failures too. Count attempts and valid lists separately. Use an initial batch to understand the variation and extend it according to the precision you need. This test does not establish a universal minimum.
  3. Measure presence, position and link. Presence is the percentage of valid lists in which the brand appears. Queseo, for example, came up in 37 of the 98 browser lists, 37.8%. Also publish how many lists place it among the top three and how many let users reach its website, and always explain the denominator.
  4. Review the content of the recommendation. Note which service is attributed, which client it is proposed for and what specific data it provides, and check important claims against their source.
  5. Record sources, listings and maps separately. A link to the candidate’s website does not, on its own, reveal where their review count came from.
  6. Compare equivalent dates and conditions. Alternate batches between environments and repeat them on different days. Any change of model or configuration must be documented.
Chapter 付録

Appendix

14 / 15 · Appendix

Full table: the 26 candidates

Presence and first positions of each candidate in the two environments, sorted by their presence in the API. A zero means they did not appear in that sample, not that they cannot appear.

CandidateBrowser listsBrowser presence1st browserAPI listsAPI presence1st API
Queseo3708787,0 %28
Jesús Porres48108282,0 %50
Alex Figueiras54256262,0 %2
Enrique Tejedo772611,0 %1
David Cerdá (Seomar)71244444,0 %6
Rocío Santamaría3504646,0 %0
Álvaro Pichó (SEO Levante)3691212,0 %1
Rafa Sospedra1404242,0 %10
Alfredo Río (SEO Local Valencia)29011,0 %0
Andrés Castillo (GEO Valencia)1632424,0 %0
Miguel Valero102626,0 %1
Paul Cris (Paul Marginean)22011,0 %0
Jamil Mansilla802323,0 %1
Paco Rubén1601717,0 %0
Jorge Segovia001111,0 %0
Softalian (Nacho Monterde)6099,0 %0
Daimako7000,0 %0
Ruk Lab7011,0 %0
Alberto Fuster2144,0 %0
agenciaSEO.eu3000,0 %0
Alban Febrer0033,0 %0
Alberto Illobre1000,0 %0
Arturo Argilés (SEO&SEM Valencia)0011,0 %0
Idital0011,0 %0
SEO Solutions0011,0 %0
Vicente Climent (Santi Pc)0011,0 %0
Bases of 98 valid lists in the browser and 100 in the API.

15 / 15 · Credits

Authorship, sources and scope

Nina S. DzhezherELNIQ SL

Eric Jorge Seguí ParejoQueseo

The study is signed by Nina S. Dzhezher, of ELNIQ SL, and Eric Jorge Seguí Parejo, of Queseo. Queseo is also one of the brands recommended in the responses analysed, which is why the same counting criterion was applied to all candidates. ELNIQ also publishes the study on its blog.

The quantitative results come from the two original Excel files of the test, one with the browser responses and one with the API responses. The external references provide methodological context and do not extend the sample:

  1. Aggarwal et al. (2024). GEO: Generative Engine Optimization.
  2. Liu, Zhang and Liang (2023). Evaluating Verifiability in Generative Search Engines.
  3. OpenAI. Web search, Sources section.
  4. NIST/SEMATECH. Confidence intervals, Wilson method.
Share the study

Want to know what AI says about your business?

Tell me what you do and which searches matter to you, and I'll tell you how to measure whether ChatGPT recommends you and what can be worked on.