GEO study: which pharma companies do AI assistants name for heart and immune conditions in the UK?
We put 20 patient-style questions about heart and immune conditions to six AI engines and counted which pharma companies they name, which sources they cite and how the company websites look to an AI crawler.
AI, TELL ME! Pharma Visibility Study, United Kingdom. Written by Natali Smith, co-founder of AI, TELL ME! Published October 2026.
Short answer. Across 120 AI answers to patient-style questions about heart and immune conditions in the UK, this GEO study found that six AI engines named 22 pharma companies 485 times, and Novartis, Eli Lilly and AstraZeneca together hold 37.7% of those mentions. Which brands appear varies by condition: AstraZeneca is named in 35 of 60 cardiology answers and in 3 of 60 immunology answers. Government bodies, journals and guideline bodies supply two thirds of the 1,040 cited items, and pharma company sites supply 9.7%. In this small sample no website property measured showed a detectable link to how often a brand is named, and the three brands named in no answer all had reachable websites in the probe.

A patient in the UK opens an AI assistant and asks which medicine, from which manufacturer, is commonly recommended for heart failure or severe psoriasis. Which companies the assistant names in that moment is what this study measures. We put 20 questions of that kind to six AI engines with a UK geo target, counted the companies the 120 answers named and the websites they pointed to, and then looked at how the companies’ own websites appear to an AI crawler.
One condition shapes everything below: the questions asked for a treatment and its manufacturer. The ranking therefore shows which companies come up when a patient asks for a manufacturer, and says nothing about how often companies come up in ordinary medical questions.
Both halves of this study ran on our own tools
AI, TELL ME! is a full-cycle AEO (answer engine optimisation) and GEO (generative engine optimisation) agency, and we operate our own monitoring and website assessment tools, which produced both halves of this study: the monitoring platform that records what each AI engine says and cites, and the AI Access Report that records what a probe using the engines’ crawler names can reach and read on a website. Every figure below traces back to the run that produced it, and the methodology section at the end documents the prompt set, the counting bases, the evidence grades and the size of the run.
The monitoring ran in English with a UK geo target:
Two segments: cardiology and immunology, with 10 prompts each.
Six engines: ChatGPT, Claude, Perplexity, Gemini, Google AI Mode and Google AI Overview, all with web search.
120 answers to patient-style questions that ask for a treatment and its manufacturer, such as which newer medication suits worsening heart failure.
Twenty-five company websites were probed once the answers were collected, 22 belonging to brands that appear in the answers and 3 to brands that never do. A brand here means a pharma company as the answers name it, such as Novartis or AstraZeneca. The study gives no medical advice.
Definitions: how to read the numbers in this study
Ten measures carry the findings below.
A mention is one count of a brand in an answer, and one answer can carry several mentions of the same brand.
Share of voice is the share of all 485 mentions that belongs to one brand. It describes conversational presence and says nothing about prescriptions or sales.
Answers naming a brand counts the answers in which the monitoring platform’s brand extraction recorded the brand at least once.
A cited item is one cited website domain in one answer, with repeats inside an answer counted (1,040 items on 229 domains), and answers citing a domain counts the distinct answers that cite it at least once.
Source type sorts every cited domain by its publisher into nine groups and an unclassified group. nih.gov counts as a government and public health body because the US National Institutes of Health runs it, although it also hosts PubMed Central full texts of journal articles.
A brand’s own site is a cited domain whose name identifies the brand.
Content extraction score is how much of a page a retrieval pipeline recovers as content once menus, filters and footers are stripped, and passage quality is how well a page survives being cut into the passages an answer is built from. Both are scores out of 100.
Structured data score is the report’s score for structure and markup: JSON-LD, machine-readable publication dates, one H1 per page, unbroken heading levels, a declared language and a page title and description.
Three brands hold more than a third of all mentions in the UK

Figure 1. Share of voice across all six engines, 485 mentions. Brands below the line sell in the UK and were named in no answer.
Novartis holds 15.7% of the 485 mentions, Eli Lilly 12.0% and AstraZeneca 10.1%, which puts the three at 37.7% between them. The ten most-mentioned brands hold 87.2% of all mentions and the other 12 named brands share the remaining 12.8%. If the 22 brands shared the mentions evenly, the top three would hold 13.6% and the top ten 45.5%.
Table 1. The ten most-mentioned pharma companies, all six engines.
Brand | Mentions | Share of voice | Answers naming the brand (of 120) |
|---|---|---|---|
Novartis | 76 | 15.7% | 54 |
Eli Lilly | 58 | 12.0% | 44 |
AstraZeneca | 49 | 10.1% | 38 |
Amgen | 39 | 8.0% | 32 |
Bayer | 36 | 7.4% | 27 |
Pfizer | 36 | 7.4% | 28 |
Sanofi | 35 | 7.2% | 33 |
Boehringer Ingelheim | 33 | 6.8% | 30 |
Merck & Co. (MSD) | 31 | 6.4% | 21 |
Roche | 30 | 6.2% | 20 |
Below the top three the order depends on how you count
Counted by answers, Sanofi ranks fourth with 33 answers, against seventh by mentions (35), and Bayer ranks eighth with 27 answers, against a tied fifth place with Pfizer by mentions (36). Novartis, Eli Lilly and AstraZeneca hold the top three places on both counts.
Nine brands appear in three answers or fewer
Biogen and Johnson & Johnson (Janssen) are named in 3 answers each; Celltrion, UCB, AbbVie and GSK in 2 each; Gilead Sciences, Kyowa Kirin and Takeda in 1 each. The nine hold 22 mentions (4.5%) between them, and one appearance in 120 answers is too little to place a brand in the ranking.
The condition changes which brands appear

Figure 2. Answers naming each brand, out of 60 answers per segment: cardiology on the left, immunology on the right.
AstraZeneca is named in 35 of 60 cardiology answers and in 3 of 60 immunology answers; Boehringer Ingelheim in 29 and 1; Bayer in 27 and 0; Merck & Co. (MSD) in 18 and 3; Roche in 3 and 17 (Figure 2 and Table 2). Novartis is fourth in cardiology (25 of 60) and first in immunology (29 of 60).
Table 2. Answers naming each brand, by segment (60 answers per segment).
Brand | Cardiology | Immunology |
|---|---|---|
AstraZeneca | 35 | 3 |
Boehringer Ingelheim | 29 | 1 |
Bayer | 27 | 0 |
Novartis | 25 | 29 |
Eli Lilly | 21 | 23 |
Merck & Co. (MSD) | 18 | 3 |
Amgen | 14 | 18 |
Sanofi | 12 | 21 |
Novo Nordisk | 11 | 0 |
Pfizer | 10 | 18 |
Roche | 3 | 17 |
Five of the ten brands named in the most answers show a segment difference that survives a correction for the number of tests: AstraZeneca, Boehringer Ingelheim, Bayer, Merck & Co. (MSD) and Roche. Novo Nordisk, which sits outside those ten, passes a separate test of similar strength. Novartis, Eli Lilly, Amgen, Sanofi and Pfizer show no such difference, and Novartis and Eli Lilly are the only brands named in at least 20 of 60 answers in both segments. These are presence rates inside a segment and carry no ranking.
For a brand team, that makes a single overall visibility figure a poor summary, because it would hide a split as wide as AstraZeneca’s 35 and 3.
Twenty-one of 120 answers name no brand
Twenty-one of the 120 answers (17.5%) name no brand, 8 of 60 in cardiology and 13 of 60 in immunology, and five of the 13 immunology answers belong to one prompt, the question about Crohn’s disease with frequent relapses. The other 99 answers name between 1 and 10 brands.
The six engines name the same leaders and differ in breadth

Figure 3. Share of each engine’s brand mentions per brand, in overall order. Shade runs from white (0%) to black (30% or more) and a zero is printed as 0.0.
The engines rank the brands alike
The six engines put the brands in nearly the same order: across the 22 named brands, the rank correlations between their mention counts run from 0.74 (Perplexity with Google AI Overview) to 0.93 (ChatGPT with Gemini), all 15 pairs at p < 0.001, and 11 brands are named on all six engines, among them the ten brands in Table 1. Novartis is first or tied first on every engine. Four brands are named on a single engine (AbbVie on Google AI Mode, Gilead Sciences on Claude, Kyowa Kirin on Perplexity and Takeda on Google AI Overview) and four more on two.
Breadth differs from 12 to 19 brands per engine
Perplexity names 12 brands in 38 mentions and its three most-mentioned brands hold 57.9% of them, while Claude names 19 brands in 106 mentions and its three most-mentioned hold 31.1% (Table 3). ChatGPT names 13 brands and makes the most mentions, 112.
Table 3. Engine profile: mentions, brands named, the share held by each engine’s three most-mentioned brands and distinct domains cited.
Engine | Mentions | Brands named | Three most-mentioned brands combined | Distinct domains cited |
|---|---|---|---|---|
ChatGPT | 112 | 13 | 42.0% | 35 |
Claude | 106 | 19 | 31.1% | 155 |
Perplexity | 38 | 12 | 57.9% | 51 |
Gemini | 56 | 14 | 42.9% | 46 |
Google AI Mode | 92 | 17 | 35.9% | 1 |
Google AI Overview | 81 | 16 | 42.0% | none returned |
All engines | 485 | 22 | 37.7% | 229 |
Mid-table brands swing by engine
Roche holds 15.8% of Perplexity’s mentions and 1.8% of Gemini’s; Sanofi 15.8% of Perplexity’s and 4.9% of Google AI Overview’s; Bayer 10.7% of Gemini’s and 4.3% of Google AI Mode’s; Novo Nordisk none of Perplexity’s 38 mentions and 4.7% of Claude’s. The leaders move less: Eli Lilly holds between 7.9% and 16.1% of the mentions on every engine and Novartis between 12.0% and 26.3%. One mention moves a brand’s share by 0.9 to 2.6 points, depending on the engine, and a brand that checks a single engine sees one sixth of the answers in this study.
Public bodies, journals and guidelines supply two thirds of the cited items

Figure 4. Cited items by source type, 1,040 items from 229 domains. The sites whose name identifies one of the 22 brands are part of the pharma company row and are not added to it.
Government and public health bodies supply 341 of the 1,040 cited items (32.8%), journals and medical publishers 227 (21.8%) and guideline, professional and patient bodies 127 (12.2%), which puts the three source types at 695 items (66.8%) between them (Figure 4).
Brand sites supply a small share of citations
Pharma company sites supply 101 cited items (9.7%) across 46 domains, and 58 of those items (5.6%) sit on 20 domains whose name identifies one of the 22 brands. By answers, 24 cite at least one of those brand-named sites, which is 30.0% of the 80 answers that cite any source. Seventeen domains carrying 33 items (3.2%) could not be typed from their names.
One domain supplies nearly a fifth of all citations

Figure 5. The 15 most-cited domains (A) and the number of domains by how many answers cite them (B).
nih.gov supplies 196 cited items (18.8%) and is cited in 46 of the 120 answers (38.3%). nice.org.uk comes second with 82 items (7.9%) in 22 answers (18.3%), and the 15 most-cited domains together supply 525 of the 1,040 items (50.5%).
At the other end, 162 of the 229 domains (70.7%) are cited in one answer only, and four (nih.gov, nice.org.uk, doi.org and springer.com) are cited in 10 or more answers. Of nih.gov’s 114 cited addresses, 84 are PubMed Central or PubMed pages, 20 are DailyMed drug labels and 10 are NCBI Bookshelf pages.
The most-cited source differs by engine
In the data this run collected, four engines came back with source links: Claude cites 155 distinct domains, Perplexity 51, Gemini 46 and ChatGPT 35. Google AI Mode returned one source (google.com) and Google AI Overview none, so the citation figures rest on four engines and say nothing about how the two Google engines source their answers. nih.gov leads Perplexity (104 of its 196 items) and Claude (67), nice.org.uk leads ChatGPT (43) and patsnap.com leads Gemini (13), so more than half of nih.gov’s citations come from one engine. Forty of the 120 answers carry no cited domain.
Most company websites served AI crawler requests, and structured data was the lowest page-based score
The 25 sites appear as Site A to Site Y, with Sites A to V belonging to the 22 brands that appear in the answers and Sites W to Y to the 3 that never do.

Figure 6. Readiness scores for Sites A to Y, one probe each. Shade runs from white (0) to black (100) and not-measured cells are hatched.
Crawler replies are reported as counts
Sixteen of the 25 sites served all 14 fetching crawler identities (Table 4). Six sites refused both plain control requests, and five of those six returned no usable result for any request, so the probe withholds any verdict about AI crawlers for them. The sixth served four named crawlers and refused ten, an allow list by name. One further site refused 14 crawler names and served the probe’s own name from the same address, which marks its refusal as an identity check.
Table 4. Crawler replies by site, as counts (25 sites, 14 fetching crawler identities each). A site that answered every request with HTTP 200, plain controls included, returned a challenge page and counts as withheld.
Reply pattern | Sites |
|---|---|
Served all 14 identities | 16 |
Refused every request or returned a challenge page (withheld) | 5 |
Served 4 named crawlers and refused 10 | 1 |
Refused all 14 crawler names and served the probe’s own name | 1 |
Served 13 identities, with 1 rate-limited (HTTP 429) | 1 |
Served 12 identities, with 2 probes that did not land | 1 |
Three of the five withheld sites belong to brands that the answers name and whose own domains the engines cite, so a statement that those sites are closed to AI sits against the monitoring data, and for the other two the contrary evidence is weaker. A refusal in a probe describes what a test client received from one address on one day and settles nothing about what a live crawler receives, and server logs are the only place to see that.
Access and text sit near the ceiling, and structured data is the lowest page-based score
Crawler access and text without JavaScript sit near the ceiling, with medians of 100 and 99.5, and structured data is the lowest of the five page-based scores, with a median of 73.5 (Table 5). Four sites score 85 or more on structure (85, 88, 89 and 100) and three score below 50 (21, 28 and 46), so the weakness is common and has exceptions. In plain terms, in the probe the sites mostly answered crawler requests and returned readable text, and the weak spot was the markup that tells a machine what a page contains. The lab speed score does not depend on page reads and was measured on 10 sites, with a median of 53.5, and the field data band reads average at seven of them and slow at three.
Table 5. Score ranges across the sites measured. Sites with no score for a measure are left out of every range and median, and Site S, read on one page, is left out of the page-based ranges and counted in the speed row.
Score | Sites measured | Range | Median |
|---|---|---|---|
Crawler access | 18 | 98 to 100 | 100 |
Text without JavaScript | 18 | 50 to 100 | 99.5 |
Content extraction | 18 | 56 to 98 | 88.5 |
Passage quality | 18 | 76 to 97 | 88.5 |
Structured data | 18 | 21 to 100 | 73.5 |
Speed (lab) | 10 | 47 to 83 | 53.5 |
Page signals differ between sites
Across the 18 sites read on 21 pages:
JSON-LD sits on every sampled page at 6 sites, on part of the pages at 7 and on none at 5.
A machine-readable date sits on every sampled page at 4 sites and on none at 5, and the two signals do not travel together: two of the six sites with JSON-LD on every sampled page carry a date on 7 and 11 of 21 pages.
The extractor keeps between 15% and 87% of the visible text (median 45%).
Across all 25 sites, an llms.txt file was verified present at 3, verified absent at 16 and could not be verified at 6, because those six sites refused plain requests.
Three brands appear in no answer, and all three sites were reachable
Servier, Menarini Group and medac sell in the UK and were named in none of the 120 answers. Their sites were reachable in the probe: each served all 14 fetching crawler identities and scores 100 on crawler access and 100 on text without JavaScript, so access gives no reason for their absence. Two of the three score 56 and 63 on content extraction, below every one of the 15 named-brand sites read on 21 pages (lowest 76), and the structure scores run from 21 (the lowest of the 18 sites) to 78 and 100 (the highest). The extractor keeps 15%, 20% and 37% of the visible text against a median of 50% at those 15 named-brand sites. These are three sites, and the pattern describes those three.
In this small sample, no measured site property shows a detectable link to how often a brand is named
We set 14 site measures, from the six scores in Table 5 to page signals such as JSON-LD and machine-readable dates, against two outcomes: how many mentions a brand receives and how many answers name it. That gives 28 tests, each a Spearman rank correlation.
Among the named-brand sites (15 per page-based test, 9 for speed) no test of the 28 reaches p < 0.05; the largest coefficient is 0.25 in absolute value and the smallest p-value is 0.37. Adding the three never-named sites at zero mentions lifts one test over the line (the share of visible text the extractor keeps against answers naming the brand: rho 0.48, p = 0.045), and the same measure against mentions gives p = 0.052. At 0.05, 28 tests would produce about 1.4 passes by chance, so one pass carries little weight.
The six sites that refused plain requests and the site read on a single page have no page-based scores and are not in the table, which covers 15 of the 22 named brands. With 15 to 18 sites no coefficient supports a claim: the sample neither confirms a link nor rules one out. The sample sizes per test and the corrected threshold are in the methodology section.
The brands named most often are also the brands whose own sites are cited
The ten most-mentioned brands each have their own site cited in the answers, from 1 to 15 items, and three of the other twelve do (Celltrion, UCB and AbbVie, 1 item each). Across the 22 named brands, own-site citations track mentions (Spearman rank correlation 0.71, p = 0.0002), and the relation holds without the leading brand (0.68, n = 21, p = 0.0007).
The ratio varies: own-site citations per 100 mentions run from 1.7 for Eli Lilly (1 item in 58 mentions) and 2.6 for Amgen to 41.7 for Bayer (15 in 36) and 50.0 for AbbVie (1 in 2).
What the results mean for a pharma company
The results point to four checks a pharma company can run on its own category before it commissions any optimisation:
Check visibility for each condition separately, because AstraZeneca is named in 35 of 60 cardiology answers and in 3 of 60 immunology answers, and a single overall figure hides that split.
Compare the engines one by one, because Roche holds 15.8% of Perplexity’s mentions and 1.8% of Gemini’s, and a brand that checks a single engine sees one sixth of the answers in this study.
Look at the pages the answers cite in your own condition area, because pharma company sites supply 9.7% of the cited items while public bodies, journals and guideline bodies supply two thirds.
Repeat the measurement on a fixed set of questions, because this study is a single run with 10 prompts per segment and a change in share of voice only counts when the questions have stayed the same.
A website audit comes after those four checks, because in this small sample no website property measured showed a detectable link to how often a brand is named.
After an audit, the work moves from measurement to a plan
A study of this kind describes a starting position, and for a single brand AI, TELL ME! runs the work that follows:
AEO Monitoring shows whether and where a brand appears in AI-generated answers, how it is described and which sources shape each response.
The GEO Report evaluates how the brand’s website, content and external presence signal to the engines what the brand is and whether it should be recommended, and identifies which signals are missing or underrepresented.
The AEO & GEO Roadmap turns the findings into a prioritised 90-day plan covering technical structure, content, messaging and external presence.
Implementation carries the plan out alongside the brand’s own team, as a fully outsourced function or as a blend of both, covering technical fixes, content, external authority building and ongoing optimisation.
Monitoring then repeats on the same prompt set, so a change in share of voice or in citations can be read against the same questions.
The data do not support copying the leaders
The visible brands have their own sites cited often, which invites a brand to copy what the leaders publish. Three findings argue against treating that as a recipe:
Brand-named sites supply 5.6% of the cited items, so most of the sourcing sits on public bodies, journals and guideline bodies that a brand does not control.
No measured site property shows a detectable link to how often a brand is named, and the three never-named sites were reachable in the probe, so matching a leader’s technical profile carries no demonstrated advantage.
A leader’s cited site may follow from its visibility as easily as cause it, because the own-site correlation has no known direction.
How other published research compares
Five published analyses bear on what this study found: two broadly agree with it, one diverges and two cannot be compared on the same basis.
Nguyen et al. audited the sources that ChatGPT, Perplexity and Google AI Overview cite for mental health questions (1,140 responses and 15,942 citations, a preprint of 31 August 2026). In English, government and public sources supply 22.6% of the citations, academic and journal sources 21.6% and the top ten domains 43.6%. This study finds 32.8% for government and public health bodies, 21.8% for journals and medical publishers and 43.9% for the top ten domains (457 of 1,040 items). The topic and the engines differ, and the paper classes PubMed, PMC and NCBI as academic where this study classes nih.gov as a government body, so the two government shares are not like-for-like.
LLM Pulse, an AI visibility vendor, analysed 824,997 citations to generic health questions in the United States across five engines and reported on 17 August 2026 that government and institutional sources supply 3.4% and academic and journal sources 8.3%, with a single PubMed Central host as the top domain at 4.08%. That diverges from the 66.8% here. The questions, the country and the classification differ (the vendor counts NCBI hosts as academic), and one host’s share is a different base from nih.gov’s 18.8% across all its hosts.
Grossman et al. compared Google results, AI Overviews and Gemini 2.5 Flash using a public benchmark of 11,500 user queries (arXiv, 30 April 2026) and found little overlap between the URLs each retrieves: an average Jaccard similarity of 0.18 between AI Overviews and organic results, 0.11 between AI Overviews and Gemini and 0.16 between organic results and Gemini. That matches the split by engine here, where nih.gov leads Perplexity and Claude, nice.org.uk leads ChatGPT and patsnap.com leads Gemini. The paper also reports that sites blocking Google-Extended in robots.txt are less likely to be retrieved by AI Overviews. This study cannot test that result, because none of the robots.txt files the probe could read disallowed any of the 16 crawler identities.
Żatuchin measured brand recommendations by three language models across five industries (50 brands, 250 queries and 3,750 responses; arXiv, revised 23 September 2026) and reports a mean top-three brand share of 48.6% and the same top brand on all three models in 104 of 250 queries (41.6%). Here the top three hold 37.7% across 22 brands, where an even split gives 13.6%, and the paper’s industries have 10 brands each, where an even split gives 30%, so the two concentration figures sit on different bases. The paper’s agreement figure counts queries on which the models share a top brand, and this study compares engine totals, in which Novartis is first or tied first on all six engines. The author is affiliated with Rankfor.AI, a vendor of AI brand intelligence tools, and the figures are read from the first version of the paper.
SE Ranking, an SEO software vendor, analysed nearly 300,000 domains (blog post of 7 November 2025), found an llms.txt file on 10.13% of them and found no effect of the file on how often a domain is cited. Here an llms.txt file is verified present at 3 of 25 sites, too few to test, and the SE Ranking result gives no reason to expect an effect.
Frequently asked questions
Which pharma companies do AI assistants name most often?
Novartis (15.7% of mentions), Eli Lilly (12.0%) and AstraZeneca (10.1%) are named most often in this test of 120 UK answers on heart and immune conditions, with Amgen fourth (8.0%). The three hold 37.7% of the 485 mentions, and the ten most-mentioned brands hold 87.2%.
Do AI assistants name different pharma companies for cardiology and immunology?
Yes: AstraZeneca is named in 35 of 60 cardiology answers and in 3 of 60 immunology answers, and Roche in 3 and 17. Novartis and Eli Lilly are the only brands named in at least 20 of 60 answers in both segments.
Which sources do AI assistants cite for medicines?
AI assistants cite mostly public bodies, journals and guideline bodies: government and public health bodies supply 32.8% of the 1,040 cited items, journals and medical publishers 21.8% and guideline, professional and patient bodies 12.2%. nih.gov alone supplies 18.8%, and pharma company sites supply 9.7%.
Do all AI engines name the same pharma companies?
The six engines name the same leaders and differ in breadth: the rank correlations between engines run from 0.74 to 0.93, and the engines name from 12 brands (Perplexity) to 19 (Claude). Roche holds 15.8% of Perplexity’s mentions and 1.8% of Gemini’s.
Does a better-built website make a pharma company more visible?
In this small sample no link could be detected: none of the 28 tests of site properties against how often a brand is named reaches p < 0.05 among the named-brand sites (15 sites per test, 9 for speed), and the three brands named in no answer all had reachable sites in the probe.
Methodology and limitations
This study in numbers: one category (pharma companies for heart and immune conditions) in one market (the United Kingdom), two segments, 20 prompts, six AI engines, 120 answers that between them name 22 brands, 485 mentions, 1,040 cited items on 229 domains and 753 cited URLs, plus 25 company websites probed. Citation figures cover the 80 of 120 answers that carried sources, and three brands that sell in the UK and that no answer names were added to the brand list at zero.
Ten prompts per segment is a test scale, and working monitoring runs 50 prompts per segment and monitors every segment on its own, because presence differs between segments (Table 2). The run is a single reading, run-to-run variance is a property of these systems and single-run figures are directional. The prompts are patient-style questions that ask for treatments and manufacturers, so the ranking describes this prompt set and another wording could change it.
Share of voice divides each brand’s mentions by the 485 mentions in all answers, or by one engine’s mentions in the engine view. Answers naming a brand are counted out of 120 (60 within a segment), and a citation percentage divides by the 1,040 cited items unless the text names an engine or a source type as its base. Brand counts follow the monitoring platform’s brand extraction, and close pairs are read as level. All 22 brands are counted from the same extraction of the same 120 answers.
The segment comparison is a chi-square test on the ten brands named in the most answers with a Bonferroni threshold of 0.005, which AstraZeneca, Boehringer Ingelheim and Bayer (each p < 0.0001), Merck & Co. (MSD) (p = 0.0003) and Roche (p = 0.0006) pass and Sanofi (p = 0.066) and Pfizer (p = 0.084) do not. Novo Nordisk sits outside those ten and was tested after the fact with Fisher’s exact test, cardiology against immunology, which gives p = 0.0006. The rank correlations between engines are Spearman coefficients across the 22 named brands.
The 14 website measures in the 28 correlation tests are the six scores in Table 5, the share of visible text the extractor keeps, four passage-quality shares (self-sufficient passages, paragraph density, section structure and lead answer in the title) and three page-signal shares (JSON-LD, machine-readable date and exactly one H1). Among the named-brand sites n is 15 (9 for speed), and with the three never-named sites added n is 18 (10 for speed). At 0.05, 28 tests give 1.4 expected by chance, and no test survives a Bonferroni threshold of 0.0018 in either set.
A cited domain is a brand’s own site when its name identifies the brand (20 domains belonging to 13 of the 22 brands), and product-brand sites whose domain does not name the company are not mapped. Counted as domain and answer pairs, pharma company sites account for 80 of 506 (15.8%) and brand-named sites for 46 (9.1%). Source types are our classification from the domain name and the publisher.
Each of the 25 company sites received one AI Access Report: 14 fetching crawler identities (8 citation crawlers and 6 training crawlers) plus two permission tokens that fetch nothing, sent to the home page from one address and compared with two plain control requests, plus up to 20 pages drawn from the sitemap across its page templates, three of them opened in a headless browser. Eighteen sites were read on 21 pages and one on a single page, and the other six have no page-based scores. The probes ran on global corporate domains after the answers were collected, with no UK sites probed. Each page-signal share describes the sampled pages only. Scores are a relative scale for tracking change over time and have not been calibrated against how often a site is cited.
Access findings carry one of three evidence grades: site-attributable (the evidence sits in a page’s HTML or a static file, or a server answered some crawler names and refused others inside one run from one address), withheld (every request was refused or returned a challenge page, so no verdict about AI crawlers is possible) and contradicted (the finding sits against other evidence in the data). Twenty sites are site-attributable and five are withheld, and a statement that a withheld site is closed to AI crawlers would be graded contradicted for three of the five, with weaker contrary evidence for the other two.
Site-level findings are published as Site A to Site Y. The letters were assigned by a seeded shuffle, and no letter is tied to a brand, a share or a rank. Each company can request its own AI Access Report.
Brands are spelt as each company writes them on its own website. Merck & Co. is known as MSD outside the United States and Canada and counts as one brand (no answer refers to Merck KGaA), and the Janssen pharmaceutical business of Johnson & Johnson appears as Johnson & Johnson (Janssen). medac is written in lower case. The study date is October 2026.
Running a GEO study for your own category
This study covers pharma companies for heart and immune conditions in the UK, and the same measurement runs for any category, in any market and language, on the same two halves: what the engines say and cite, and what a probe using their crawler names can reach and read.
We bring 17 years in marketing, PR and SEO to this work.
Book a 30-minute review of your own category at https://calendly.com/natali-khld/welcome.
Request the prompt set, the engine-by-engine breakdown or the AI Access Report for your own website at hey@aipleasetellme.com.
Written by Natali Smith, co-founder of AI, TELL ME! (aipleasetellme.com), October 2026.


