What a Uniform Supplier and a Bank Have In Common, According To AI

Same measurement, two unrelated industries, three real patterns survive the comparison

SiteSignal has now run full AI-visibility studies on two categories that could hardly be more different: a fragmented, SME-heavy UK apparel market, and a tightly regulated, six-bank Sri Lankan oligopoly. Comparing them directly separates what's true about AI visibility everywhere from what's just true about workwear, or just true about banking.

• Contact details are the least trustworthy fact type in both sectors, by a wide margin, and identity basics (name, founding date, HQ) are the most trustworthy in both. This is the strongest "this generalizes" finding across the two studies (Section 05).
• AI almost never cites a negative source about anyone, in either sector, under 1% negative-sentiment citations in both. Two unrelated categories, the same near-total absence of visible criticism (Section 03).
• Market structure predicts leaderboard stability almost perfectly. Banking's six licensed majors held their positions through a full platform migration that scrambled workwear's leaderboard almost completely. A regulated oligopoly is robust to how you ask the question; a fragmented market is not (Section 02).
• The two sectors diverge sharply on where AI gets its information, workwear's citations shift steadily away from brand-owned sites over time; banking's don't move at all. Workwear picked up a Reddit citation explosion; banking picked up almost no social-platform citation activity at all (Section 04).
• Both platforms commit to more confident answers about banks than about workwear suppliers, plausibly because banks publish more structured, checkable public information. If that holds beyond these two studies, information density, not category prestige, is what earns a model's confidence (Section 06).

A long tail versus a settled six

The two markets differ in shape before a single AI answer is even considered, and that shape alone predicts most of what follows.

• 0.22% of the workwear directory forms its stable core; 3.4% of the banking directory forms its stable core, a market that's over fifteen times more concentrated, proportionally, at the top.
• Both studies span the same May 2026 monitoring-platform migration, which rebuilt the tracked prompt set from scratch. Workwear's top-ten turned over almost completely across that seam, nine of ten June leaderboard names hadn't appeared in April's top ten. Banking's top six were identical before and after.
• The honest reading isn't "banking's AI visibility is more stable", it's that a market with six licensed, heavily-documented competitors has one obvious answer to almost any question you ask about it, so the exact wording of the question matters less. A market with hundreds of similarly-sized SME suppliers doesn't have that anchor, so which specific names surface depends much more on which specific questions get asked.

One line moves. One line doesn't.

• Workwear's owned-site share fell 38 points across eight tracked months. Banking's moved 4 points, with no clear direction. The same monitoring methodology, applied to two different markets, found a real trend in one and genuine stability in the other, which is itself evidence the workwear trend is a real market shift, not a measurement artifact.
• The likely reason is structural: a uniform supplier competes against independent buying guides, comparison blogs and forum threads that don't exist in the same volume for a licensed bank's product line, banking has fewer independent substitutes for "what does this account actually cost," so the bank's own site stays the default source.
• Reddit went from invisible to over one in nine workwear citations inside ten weeks. Nothing comparable happened in banking, its highest social-platform share in the entire window was barely 1%. Whether people turn to social platforms for opinion on a purchase category looks like a real, measurable, category-specific behavior, not something AI visibility monitoring creates on its own.

A cliff in one category, a slope in the other

• Both categories show the same direction, more specific questions name more specific brands, but workwear's gap is nearly two and a half times wider than banking's.
• Banking starts from a much higher awareness-stage floor (51.5% vs 25.2%): even a generic question like "what is a fixed deposit" tends to get answered with reference to a specific bank's actual product, because rates and terms are inherently bank-specific information rather than generic category knowledge.
• Practical read for either kind of business: in a category with workwear's shape, category-education content is close to wasted on AI visibility, the decision-stage query is where the opportunity lives. In a category with banking's shape, education-stage content already carries real weight, because the products themselves are specific enough that "explaining the category" and "naming a provider" aren't fully separable questions.

The one pattern that looks like a rule, not a coincidence

This is the strongest cross-sector finding in either study. Two industries, two continents, two completely different fact sets, and the same field comes out least reliable both times.

*Workwear's business-name field was directly comparable; banking's equivalent check was excluded from its own report after inspection showed it flagging correct legal-suffix answers ("Bank PLC") as wrong against an incomplete reference record, a ground-truth data issue, not a hallucination, so it isn't used in this comparison either. †Banking's Terms URL (16.7% match) and Privacy Policy URL (54.2% match) diverged from each other more than workwear's did; see that report's own note on banks restructuring these pages often enough to confuse a static reference.

• Contact-detail fields posted the single worst mismatch rate in both reports independently, 30.0% for workwear's support email, 56.9% for banking's public contact email. Neither report was built expecting to find the other's result.
• Sentiment on citations was likewise near-identical: 42.8% positive / 56.7% neutral / 0.5% negative (workwear) against 46.1% / 53.7% / 0.2% (banking). AI essentially never cites a source to say something unflattering about anyone, in either market.

Both models are more confident about banks

• Perplexity hedges 9 points less often on banking questions than on workwear ones. ChatGPT's hedge rate is already low in both, but still lower for banking.
• The likely driver isn't the platform, it's the training data. Banks publish extensive, structured, regularly-updated public information (rate sheets, regulatory disclosures, annual reports); most SME uniform suppliers don't have an equivalent volume of citable, structured public content. Less to work with plausibly means more hedging, independent of which model is doing the answering.
• One place the pattern doesn't hold: ChatGPT's raw match rate (right, not just confident) was actually slightly higher for workwear than banking, commitment and correctness aren't the same axis, and banking's higher commitment doesn't automatically mean higher accuracy.
• Gemini hedges hardest in both categories (71.3% insufficient-data for workwear, 29.2% for banking on a much smaller sample), the one place platform personality looks consistent regardless of category.

The sector that needs it most is the one doing it least

• Banking's audit sample is small (3 runs against workwear's 21), but the direction is the opposite of what you'd want: the sector that leans hardest on its own website as an AI citation source is the one where none of the audited sites explicitly allowed the crawlers that feed ChatGPT, Gemini and Perplexity.
• This isn't necessarily a banking-specific failure, financial-services sites are frequently more locked-down by default (security posture, bot-mitigation vendors tuned aggressively) than an SME supplier's marketing site. But the effect is the same regardless of cause: content that's disproportionately likely to matter is disproportionately likely to be unreachable.
• The workwear number (48%) isn't a success story either, it means the coin-flip failure rate is the better of the two results found so far.

Sorting the findings into two piles

Likely universal: found independently in both, unrelated, sectors

• Contact-detail fields (email, phone, policy URLs) are the least reliable fact type an AI model will state about a business.
• Stable identity facts (name, HQ, founding date) are the most reliable.
• AI citations carry positive or neutral sentiment essentially all the time, visible negative citation is a rounding error, in both markets.
• More specific, decision-stage questions surface more specific brand names than generic, educational ones, the direction holds everywhere tested, even though the size of the effect doesn't.

Category-specific: true in one sector, not evidenced in the other

• A steady multi-month shift away from owned-site citations: found in workwear, absent in banking.
• An explosive single-platform citation surge (Reddit): found in workwear, absent in banking.
• A leaderboard that survives a full prompt-set rebuild intact: found in banking, absent in workwear.
• A wide gap between education-stage and decision-stage visibility: present in both, but nearly two and a half times larger in workwear.

Worth testing on a third, different sector

Two data points make a comparison, not a law. The strongest candidates to confirm or break here: does the "banks get more confident answers because they publish more structured data" theory hold for another information-dense, regulated category (insurance, telecoms)? Does the AI-crawler-access gap replicate in a third security-conscious sector? SiteSignal's next cross-sector study is the natural next step to find out.

Built from two independent SiteSignal studies

• Every figure in this report is drawn from SiteSignal's own published studies: The State of AI Visibility in UK Uniform & Workwear (Nov 2025–Sep 2026) and The State of AI Visibility in Sri Lankan Banking (Nov 2025–Jun 2026, with a six-week untracked gap in April). Consult either report directly for full methodology, monthly detail and complete institution directories.
• The two studies used the same metric definitions (mention, visibility share, source type, funnel stage, fact-check status) but different monitored accounts, different date ranges, and different prompt sets, differences in absolute numbers should be read with that in mind; the patterns highlighted here are the ones that held up regardless.
• No client account is named or identifiable in this report or in either underlying study.

The headline worth remembering: AI visibility isn't one undifferentiated thing that behaves the same everywhere. Some of what shapes it, market concentration, how much structured public information exists, whether a category has an active social-commentary culture, is genuinely category-specific. But underneath that, at least one property looks like a real property of how these models handle facts about any business at all: they're good at your identity, and they're bad at your contact details.