Tilly Norwood AI Visibility Study 2026

Particle6 had the highest appearance rate in this pilot study of Tilly Norwood and AI actors, appearing in 37.5% of 64 returned answers. Its 95% interval was 26.7% to 49.7%. The result is directional only: the intervals overlap across the leading brands, so this pilot cannot support a claim that Particle6 is more visible than another brand.

The study measured 16 prompts across Perplexity, ChatGPT, Gemini, and Claude. Tilly Norwood appeared in 20.3% of returned answers, while Xicoia Ltd. appeared in 25%. The main commercial risk is uneven recall: several names appeared regularly in one system and not at all in another.

These results describe the measured prompts on September 23, 2026. They do not describe all questions about AI actors, consumer demand, or the category as a whole.

Which brands appeared most often?

Particle6 ranked first by appearance rate at 37.5%, followed by Synthesia at 28.1%, Xicoia Ltd. at 25%, and Tilly Norwood at 20.3%. The appearance denominator is 64 returned answers. The ranking is a descriptive tally in a pilot, not a category ranking claim.

Among the named entities in the brief, Particle6 appeared in 24 answers, Xicoia Ltd. in 16, Tilly Norwood in 13, and SAG-AFTRA in 11. The study also surfaced tools, studios, media companies, publishers, and individuals. An appearance means the entity was named in an answer, not that the answer endorsed it.

Appearance leaderboard

Appearance rate is the share of the 64 returned AI answers in which each brand was named.

The leading intervals overlap. Particle6's 26.7% to 49.7% interval overlaps Synthesia's 18.6% to 40.1%, Xicoia Ltd.'s 16% to 36.8%, Tilly Norwood's 12.3% to 31.7%, and SAG-AFTRA's 9.9% to 28.2%. The differences are within the margin of this sample.

The top ten profiles show that appearance varied more by engine than a single category-wide rate suggests. All rates below use each engine's 16 returned answers, while owned-domain citation rates use 64 answers with retrieval evidence.

Where do the engines disagree?

The engines disagree most sharply on the entities central to the Tilly Norwood story. Perplexity, ChatGPT, Gemini, and Claude each returned 16 answers with retrieval evidence. Their differences can reflect both model behavior and the web-search tool used by each system, rather than a stable measure of any brand's visibility.

Engine-by-brand appearance

Each cell uses that engine's own answer count as its denominator, so engines can be compared row by row.

Perplexity named Particle6 in 10 of 16 answers, Xicoia Ltd. in 8, and Tilly Norwood in 5. Claude named Particle6 once and did not name Xicoia Ltd. or Tilly Norwood. Gemini named Tilly Norwood in 8 of 16 answers and Particle6 in 8, while ChatGPT named Tilly Norwood in none of its 16 answers.

ChatGPT concentrated more heavily on several production and technology companies. It named Synthesia in 7 of 16 answers, Hour One in 6, D-ID Ltd. in 5, Apple Inc. in 5, and Microsoft Corporation, Amazon.com, Inc., Soul Machines, and OpenAI, Inc. in 4 each. Perplexity did not name Hour One, Soul Machines, OpenAI, Inc., Nvidia Corporation, Microsoft Corporation, or Apple Inc. in its 16 answers.

Publisher and interview-related entities also split by system. Perplexity named The Hollywood Reporter in 4 of 16 answers and Piers Morgan Uncensored in 3. ChatGPT did not name The Hollywood Reporter and named Piers Morgan Uncensored once. Gemini and Claude named neither. These patterns do not establish a lasting advantage for any system because the study used one trial per prompt and is pilot depth.

Which question clusters carry each brand?

The cluster view identifies the brand named most often within each measured demand group. It separates general buyer-intent questions from sub-category and specific-need questions, so the groups should not be treated as one type of demand.

Cluster leaders

The same answers regrouped by the demand cluster each prompt was sampled for; cluster rates use that cluster's own answer count.

The chart shows where the measured prompts connected particular names with recognition, the Piers Morgan interview, SAG-AFTRA and stakeholder positions, commercial viability, reputation, or category comparison. It is useful for locating a brand's measured association with a topic. It does not establish sentiment, factual accuracy, or a broader audience preference.

Where do citations come from?

Owned-domain citations were uneven among citation-eligible entities. Deadline had the highest owned-domain citation rate at 18.8% of 64 answers with retrieval evidence. Synthesia and HeyGen each had 15.6%, while Tilly Norwood had 12.5%, SAG-AFTRA had 10.9%, and Particle6 had 7.8%.

Owned-domain citations

Share of the 64 answers with retrieval evidence that cited a URL on the brand's own website.

Xicoia Ltd., Netflix, Inc., Apple Inc., Warner Bros. Discovery, Inc., Paramount Global, Meta Platforms, Inc., Piers Morgan Uncensored, and several other citation-eligible entities had 0% owned-domain citation rate in the measured evidence set. That is a measured gap in this sample, not evidence that no useful owned material exists.

Publishers cited by the engines

Publishers cited as evidence, counted once per answer rather than once per link.

Wikipedia sources had the widest publisher reach, with en.wikipedia.org cited in 40.6% of the 64 answers with retrieval evidence. CBC appeared in 28.1%, CBS News in 20.3%, and Deadline and au.variety.com in 18.8% each. Among the named organizations, tillynorwood.com, synthesia.io, and sagaftra.org appeared in the source list at 12.5%, 12.5%, and 10.9% respectively.

Who is winning and who is at risk?

Measured strengths

Particle6 has the highest measured appearance rate at 37.5% and appeared in every engine, though its spread ran from 62.5% in Perplexity to 6.3% in Claude. Synthesia appeared in every engine and had a 28.1% appearance rate, with a 15.6% owned-domain citation rate. HeyGen also appeared in every engine, had a 17.2% appearance rate, an average first mention position of 4.1, and a 15.6% owned-domain citation rate.

Measured exposure risks

Tilly Norwood appeared in 20.3% of returned answers but was absent from all 16 ChatGPT answers and all 16 Claude answers. Xicoia Ltd. appeared in 25% of returned answers but was absent from Claude's 16 answers and had a 0% owned-domain citation rate. Netflix, Inc. and The Walt Disney Company each appeared in 10.9% of returned answers and were absent from Gemini and Claude, while their owned-domain citation rates were 0% and 1.6%.

What does the tally say about the category?

The category is fragmented in this pilot. Particle6 led the descriptive tally at 37.5%, yet its interval overlaps those of ranks two through five. No brand can be called more visible than those peers on this sample.

Cross-engine consistency varied. Particle6, Synthesia, HeyGen, and D-ID Ltd. each had 100% cross-engine consistency. Xicoia Ltd. had 75%, Tilly Norwood had 50%, and The Hollywood Reporter had 25%. The field therefore contains both names that travel across systems and names tied to one or two engines.

Third-party publishers carried much of the retrieved evidence. en.wikipedia.org, cbc.ca, cbsnews.com, deadline.com, au.variety.com, variety.com, youtube.com, latimes.com, and hollywoodreporter.com all appeared among the most-cited sources. Owned sites appeared too, but their rates were lower than the leading publisher rate in this measured set. The brief's named entities were not absent across the tally: Particle6, Xicoia Ltd., Tilly Norwood, and SAG-AFTRA all appeared, while their engine coverage differed.

What changed during the measurement window?

The data contains one measurement window and one trial per prompt, so it does not measure changes over time. It does show cross-engine divergence, publisher concentration, and uneven owned-domain citation within the window.

What should brands measure and build next?

• Test Tilly Norwood's absence in ChatGPT and Claude. Publish source pages that directly address the measured recognition, incident, stakeholder, viability, reputation, and comparison topics, then repeat the same 16-prompt study and check whether its 0 of 16 appearances on those systems changes.
• Test Xicoia Ltd.'s owned-domain citation gap. Build pages that state its relationship to Particle6 in plain language and cover the measured commercial and category questions, then re-measure whether its 0% owned-domain citation rate rises.
• Separate Particle6 material by question type. Particle6 appeared in 10 of 16 Perplexity answers, 8 Gemini answers, 5 ChatGPT answers, and 1 Claude answer. Re-run the prompt clusters after publishing distinct pages for the measured interview, stakeholder, viability, and comparison themes.
• Protect the Tilly Norwood source path. Tilly Norwood's domain appeared in 12.5% of answers with retrieval evidence, yet the name's average first mention position was 9.3. Re-measure first mention position after creating pages that answer one measured question each before linking to wider background.
• Compare owned content with media-source coverage. Deadline reached 18.8% of evidence-backed answers and The Hollywood Reporter reached 15.6%, while Xicoia Ltd. reached 0% through its domain. Re-run the study after any new owned pages and track owned-domain citations alongside publisher citations.
• Audit the Claude gap. Claude named Particle6 once, Synthesia three times, HeyGen three times, and did not name Tilly Norwood, Xicoia Ltd., SAG-AFTRA, Netflix, Inc., or The Walt Disney Company. Re-test the same prompt set after source changes rather than treating the Claude result as a category-wide verdict.

What does this mean for brands in the category?

Interpretation: recognition in generative answers depends on both topic association and the system answering the question. A name that appears often in Perplexity or Gemini may remain absent in ChatGPT or Claude on the same measured question set.

Interpretation: owned pages matter as evidence routes, but they are not the only route. The source tally includes major publishers, reference sources, video, social platforms, forums, and company sites. Brands can use repeat measurements to separate a name-recognition problem from an evidence-distribution problem.

How was this AI visibility study measured?

This pilot measured 16 prompts once in Perplexity, ChatGPT, Gemini, and Claude, producing 64 attempted and 64 returned answers. Every returned answer had retrieval evidence. Appearance rates use 64 returned answers as the denominator. Citation and source rates use 64 answers with retrieval evidence as the denominator.

The study is directional only. It has one trial per prompt, a limited prompt set, and overlapping appearance intervals among the leading entities. It cannot support category ranking claims or claims that one brand is more visible than another.

How this study was measured

• Sampling frame fixed first: 16 prompts allocated across the demand clusters before anything was asked.
• Same questions put to every engine: Perplexity (sonar-pro), ChatGPT (OpenAI), Gemini (Google), Claude (Anthropic), 1 trial(s) each.
• 64 answers returned of 64 attempted, 64 with retrieval evidence.
• Product lines rolled up to parent brands using a registry fixed before scoring.
• Appearance rate reported with a 95% Wilson interval; citation metrics counted only on answers with evidence.

Methodology appendix

This appendix is generated directly from the measurement record for this study. It is included so any brand named here can reproduce or challenge the result.

What was measured

• Category: Tilly Norwood & AI Actors
• Measurement date: 2026-09-23
• Depth: pilot — 16 prompts, 1 trial(s) per prompt per engine, 4 engines
• Answers attempted: 64
• Answers returned (appearance denominator): 64
• Answers with retrieval evidence (citation denominator): 64 (100% of attempted)
• Answers that failed outright: 0
• Analysis version: study-methodology-v2 · entity registry: entity-registry-2026-08-27

Claim limit: Pilot depth: 16 prompts measured once per engine. Directional only — not a basis for category ranking claims or for stating that one brand is more visible than another.

How the prompts were sampled

Fixed sampling frame: 12 buyer-intent need(s) from the client's brief are asked verbatim, one prompt each, grouped under: Recognition & Baseline Awareness / Incident-Specific / Industry Reaction & Stakeholder Positions / Business Viability & Commercial Case / Sentiment & Reputational Trajectory / Category-Defining / Comparative; then (the prompt budget of 16 could not give a dedicated cell to Recognition & Baseline Awareness, Incident-Specific, Industry Reaction & Stakeholder Positions, Business Viability & Commercial Case, Sentiment & Reputational Trajectory, Category-Defining / Comparative, so that demand is covered inside the generic clusters instead), then a documented set of generic commercial demand clusters (general best, everyday mainstream choice, price/value, primary use case, beginner, retail channel, gifting/occasion, specific need, comparison, trust, ownership attribution, quality tier, durability, specialist, emerging) receives the remaining allocation round-robin. The prompt writer supplies wording inside these cells only and does not choose the category's demand structure.

Results are not strictly comparable across engines: each engine uses a different model, a different retrieval tool and a different answer format, so a gap between two engines can reflect system design as much as brand visibility.

How each metric is defined

• Appearance rate — the share of returned answers that named the brand. Measured on every answer the engine returned, whether or not it searched the web, so an engine's retrieval behaviour does not silently change the denominator.
• Macro appearance rate — the mean of the per-prompt appearance rates, so each prompt counts once regardless of how many engines answered it.
• 95% interval — Wilson score interval on the appearance rate. Where two brands' intervals overlap, this study does not support a claim that one is more visible than the other.
• Owned-domain citation rate — the share of answers with retrieval evidence that cited a URL on the brand's own domain. This is a narrow measure of the brand's own site being used as a source; it is not a measure of total mentions.
• Share of all citations — the brand's own domain as a share of all attributed publisher citations in the study.
• Not measured — the brand's official domain could not be established, so no citation result exists. This is never reported as 0%.
• 0% — the metric could be evaluated and the event did not occur.
• Average position — mean position of the brand's first mention in the answer's ordered list of named brands.
• Cross-engine consistency — the share of engines that named the brand at least once.

Evidence rules

• An answer counts as having retrieval evidence only when the engine returned a retrieval receipt (search annotations or grounding records). A URL the model merely typed into its prose is not treated as evidence that it searched.
• Sources are attributed to the publisher behind each retrieval receipt. Intermediary and redirect hosts are never reported as publishers.
• Publisher identity could not be established for 0 retrieval receipt(s); those are disclosed here rather than attributed or dropped.
• Source counts are answer-level: a publisher counts once per answer in which it was cited, not once per URL.

Entity handling

Product lines and model names roll up to their parent brand for brand-level metrics, using a registry fixed before scoring. Nothing is merged on the basis of this study's own counts alone.

No roll-ups were applied in this study.

40 further name(s) picked up in the answers were not confidently resolvable to a brand in this category and are held for human classification before the next study; they carry no metrics here.

Limits of these findings

• This is a point-in-time measurement of what the named systems returned for these exact prompts on 2026-09-23. Engine behaviour changes without notice.
• The prompt set is a stratified sample of category demand, not a census. Results describe these prompts and should not be read as representative of all searches in the category.
• Where intervals overlap, differences between brands are not statistically distinguishable at this sample size.
• A brand not listed here was not measured as absent — it simply was not named in these answers.
• Appearance is a visibility measure, not an endorsement, a quality judgement or a market-share estimate.
• 1 of 25 ranked brand(s) had no official domain established before scoring, so they carry appearance results only and are excluded from citation metrics and source tables rather than being reported as zero.