Most common problem
64%
In 32 companies, at least one AI platform placed the business in the wrong or incomplete competitive frame.
SJK Labs
The B2B AI
Legibility Report
What Machines Tell Buyers Before You Do
How AI systems describe 50 regulated businesses.
An inaugural study in narrative legibility.
50
Companies audited
4
AI platforms tested
6
Dimensions scored
Inaugural report May 2026
What we already know
AI is already shaping how companies are understood before a human reaches the site.
Gartner reports that 67% of B2B buyers prefer a rep-free experience and 45% use AI in the process. G2 reports that 51% of buyers now start with AI more than Google. Forbes reports that 60% of purchases end with no clicks to website.
This inaugural study exists because narrative architecture is now the work of building a story that holds up under both human and machine reading. The findings are not really about what AI gets wrong, but about what happens when a company's narrative was built only for human readers.
Wrong company risk
46%
23 out of 50 companies were coded with an entity recognition, name collision, disambiguation or wrong company issue.
Cross-model volatility
52%
26 out of 50 companies had a gap of at least 10 points between their best and worst platform score.
Crawlability issue
22%
11 out of 50 companies were coded with a crawler, access, JavaScript, Cloudflare, gated site or homepage visibility issue.
What is Narrative Architecture
Narrative architecture is the underlying structure that holds a business story together.
Narrative architecture is the underlying structure that holds a business story together - the decisions about who the story is for, what's at stake, and how it sounds. It's not the words on the page. It's the logic underneath them that determines whether those words do any work.
Most businesses have content. Very few have architecture. The difference shows up when a buyer tries to explain what you do to someone else, when AI tries to describe you, when a journalist tries to place you in a story. If the architecture is missing, the story collapses under those conditions, and no amount of copywriting fixes it.
The SJK Labs Legibility Score
The scoring framework SJK Labs developed to run this study consistently across 50 companies.
The Legibility Score is the scoring framework SJK Labs developed to run this study consistently across 50 companies. It measures how accurately AI systems describe a business when asked buyer style questions, scoring each platform response across six dimensions.
Each dimension is rated 1-5, giving a maximum of 30 points per platform and 120 overall. It matters because the way AI describes your business is your narrative now, whether you wrote it or not.
If a buyer asks ChatGPT what you do and gets a wrong answer, a generic answer, or someone else's answer entirely, that's the story they're carrying into every conversation with you. You didn't write it. You can't correct it in real time. And most businesses have no idea it's happening.
The Legibility Score makes that visible. It turns an invisible problem into a measurable one.
Results
Headline problem patterns, average dimension scores and average platform scores
Source: SJK Labs audit dataset, n=50 companies. Manual problem-mode codes were derived from the audit notes.
Chart 1 • Problem patterns across 50 companies
Category fit was the most common problem pattern.
The most common score based problem was category fit. In 32 companies, at least one AI platform placed the business in the wrong or incomplete competitive frame.
Chart 2 • Average score by dimension
Proof was easiest. Buyer logic was harder.
Across this financial and regulated dataset, proof and credibility scored highest on average. The weaker dimensions were customer pain point, category fit and differentiation.
Chart 3 • Average score by platform
The same company can be legible in one system and collapse in another.
Claude was the highest-scoring platform in this dataset, largely because it was better at synthesis and at preserving commercial context. But the platform picture is uneven.
Seven things the audit shows
Seven things the audit shows
These are the patterns the audit surfaces repeatedly across the dataset. They are the parts of the story most likely to break once AI systems become the first interpreter of a business.
01
AI can often name the category, but it still loses the commercial difference.
The central problem is not total ignorance. In many cases, AI systems can identify the broad category but the problem is that the reason to buy gets compressed.
30/50 companies were coded with a narrative, category, wedge, freshness or differentiation issue
02
Name collisions are now a board level communications risk.
The most damaging problems are wrong company answers.
23/50 companies were coded with an entity recognition, name collision, disambiguation or wrong company issue
03
The same company can be legible in one AI system and invisible in another.
AI perception is a landscape. A company can score strongly in Claude and collapse in Gemini. It can be understood by Perplexity on one question and confused with a city, food brand or algorithm on the next.
26/50 had a 10-point platform gap or more
04
AI often knows the previous version of the company better than the current one.
The audit repeatedly shows that AI does not update just because a homepage changes. It updates when the new story is repeated, corroborated and indexed across the broader authority layer.
A rebrand without a machine readable cascade leaves AI telling the old story
05
Proof exists, but AI often cannot use it.
Many companies had strong proof on their websites and AI still missed it.
17/50 companies were coded with a proof, credibility or evidence surfacing issue
06
Crawlability is now a communications problem, not just a technical one.
When AI systems cannot read a company’s own website, they do not stop answering. They answer from whatever is available.
11/50 companies were coded with a crawler, access, JavaScript, Cloudflare, gated site or homepage visibility issue
07
AI is weakest when asked to compare.
The alternatives question exposed one of the clearest buyer risk patterns. When companies do not define their competitive set, AI fills the gap with scraped database logic, category adjacent guesses, outdated competitors or entirely wrong industries.
Comparison is where buyers move from curiosity to decision
Named examples
Snapshots
The full audits are the dataset. The public report uses a smaller set of named examples to make each problem mode tangible. These examples are not a ranking; they are evidence of different AI problem patterns.
Zepto
Entity collisionAll four models described the Indian quick-commerce unicorn. None identified the Australian account-to-account payments infrastructure business at zepto.com.au.
Why it matters: a clear website is not enough when a stronger namesake dominates the graph.
Shift
Generic name riskAcross 24 responses, no platform identified Shift Australia. The models defaulted to browsers, insurance AI, used cars, keyboard keys and calculator functions.
Why it matters: generic brand names without a disambiguation footprint are not reliably machine-resolvable.
Wonderful
Wrong companyAI described a Californian pistachio conglomerate, an AI agent platform, or the adjective “wonderful”. No platform found the UK Pay by Bank business.
Why it matters: if the name does not pin the company, better-known meanings will win the query.
Oakbrook
CrawlabilityThe homepage was Cloudflare-blocked, so AI reconstructed the UK fintech from third-party fragments, unrelated namesakes and criticism.
Why it matters: once first-party access fails, the market learns the company in other people’s words.
Ferovinum
Freshness lagModels understood the funded wine and spirits fintech, but they were one rebrand behind: Fero, ferodrinks.com, FLOW and updated metrics were largely invisible.
Why it matters: rebrands need a machine-readable cascade, not just a new homepage.
Payhawk
Narrative wedgeAI recognised enterprise spend management, but often missed the key wedge: Payhawk’s ability to manage existing corporate card programs rather than forcing migration.
Why it matters: a company can be visible and still have its best commercial difference flattened out.
Hippo Insurance
Outdated storyMost models still described the legacy smart-home carrier story, while the current site presents a 70+ carrier quote-comparison agency.
Why it matters: AI often remembers the old strategic form of the company better than the current one.
Tokenovate
Technical flatteningAI captured post-trade automation and CDM, but often flattened the edge: nine CDM-native workflows, the Unified Trade Record and Novat as settlement-act tokenisation.
Why it matters: deep technical differentiation is often the first thing to disappear under summary pressure.
Coinbase
First-party accessAI could describe the product surface accurately, but it could not read the homepage. One platform even made a false security claim that a first-party incident page could have corrected.
Why it matters: if the canonical narrative is unreadable, secondary sources become the de facto brand voice.
AlphaSense
Proof underusedModels captured the category and scale, but flattened the proprietary content moat: Tegus, Wall Street Insights, no-hallucination architecture and Gartner MQ proof were underused.
Why it matters: even well-covered companies can lose the proof that actually supports premium positioning.
Moula
Disambiguation driftWhen the model found the lender, it often described the category correctly. On softer questions, some systems drifted into the Arabic term “Mawla” or older loan ceilings.
Why it matters: even partial recognition can break once the question becomes less anchored and the name has stronger alternate meanings.
R3
Blocked homepageBecause r3.com was effectively unreadable to non-JavaScript retrieval, one platform reconstructed the enterprise blockchain company, another asked for clarification, and one thought “R3” meant a game controller button.
Why it matters: if the first-party site is inaccessible and the name is overloaded, AI fills the gap with whatever surface is easiest to resolve.
Full scoring table
Full scoring table
Companies are sorted by average total score across the four platforms. Each platform total is out of 30.
View the full scoring table
The separate scoring-table page includes the full ranking table from the report itself, including ChatGPT, Claude, Gemini and Perplexity scores, plus average score and spread.
Methodology
How the audit was run
The study measured AI generated explanations of companies. It did not measure actual buyer behaviour, conversion, search ranking, website traffic, media coverage, analyst opinion or commercial performance.
Step 01
Company selection
The study audited 50 regulated, financial, fintech, insurance, wealth, payments, crypto and capital intensive businesses. Companies were selected to represent a range of visibility levels, naming risk, technical complexity and narrative maturity.
Step 02
Platforms tested
Each company was tested across four web-search-enabled AI platforms: ChatGPT, Claude, Gemini and Perplexity. The study does not attempt to create a sterile model benchmark; it measures how a buyer, journalist, analyst or partner would experience each platform in normal use.
Step 03
Prompts used
Each company was tested using the same six buyer style questions: what does this company do, who is it for, what problem does it solve, what makes it different, is it credible, and who are the main alternatives?
Step 04
Scoring framework
Responses were scored using the SJK Labs Legibility Score across six dimensions: clarity, accuracy, differentiation, customer pain point, proof and credibility, and category fit. Each dimension was scored from 1 to 5, giving each platform a maximum of 30 points per company.
Step 05
What the study measured
In several cases, the AI’s answer was not simply incomplete; it returned the wrong company, the wrong category, stale positioning, irrelevant competitors or third-party criticism without first party counterweight.
Step 06
Treatment of first-party access
For each company, first party site content was treated as the primary ground truth where available. Where the company website was blocked, JavaScript-walled, cookie-walled, Cloudflare-challenged, security-gated, redirect-only or effectively empty, that inaccessibility was recorded as part of the finding.
What companies should do now
What companies should do now
The practical lesson is not to write more content, but to make the company easier for AI systems to explain accurately.
Companies should treat AI legibility as a distinct layer of communications strategy - part positioning, part PR, part technical visibility, part proof architecture.
About SJK Labs
For businesses that know the work is strong, but suspect the market signal is weaker than it should be.
SJK Labs helps businesses turn expertise into authority infrastructure: clearer positioning, stronger proof, sharper media logic and a digital ecosystem that can hold up in an AI-first world.
If buyers, journalists, search engines and AI systems cannot clearly understand what makes you credible, someone less capable but more legible will fill the gap.
Want to see how legible your own business looks in AI?
The Legibility Audit measures how clearly AI systems can explain your business, where they lose the commercial difference, and what story a buyer is likely to carry into the next conversation.