We Ran the FTSE 100 Through the Crane Index^™. Most of It Could Not Be Read.

60 of the FTSE 100 could not be read by a machine at all

On 6 August 2026 we ran the FTSE 100 through the Crane Index™, our three-scanner AI-visibility suite. 21 of the 100 constituents do not resolve to a single primary corporate domain a machine could even approach, which is an AI-legibility failure in itself. Of the 79 that do, 39 could not be scored at all: 23 refused our automated reader outright, 9 gave it no response, and 7 serve pages that do not exist until scripts execute, which is its own kind of illegibility. Of the 40 sites a machine could actually read, the median AI Search Readiness score was 47 out of 100, and not one reached the Strong band of 70 on the Crane Index. These are the largest, best-funded companies in the country, and budget bought none of them legibility. All figures are first-party measurements; the full, dated result set is archived, and we will re-run the scan and report the movement.

Budget does not buy machine legibility

The prevailing view, sold by technical SEO agencies and reflected in every marketing dashboard, is that a site with healthy traffic, strong brand search and solid rankings is fundamentally sound. For AI search that view is inverted. Traffic measures how humans reach a site. AI readiness measures whether a machine can resolve the company as a single entity and repeat its commercial facts without error. Those are different tests, and the estates in our scan mostly pass the first and fail the second, because a large enterprise site is usually a federation of content management systems, migrations, agency handovers and template debt that no one owns end to end. Every seam loses a signal a machine relies on. A ten-page challenger with clean structured data can outscore a FTSE 100 site with ten thousand pages, because page count does not aid extraction, it dilutes it, and this scan bears that out: what separated the top of the field from the bottom was structure, not spend.

A practitioner will object that this is a JSON-LD linter with a threshold. The score is broader than that, but the objection points at the right variable. The strongest sites we measured carry valid Organization structured data. The weakest do not. Entity disambiguation, the sameAs links that tell a machine which real-world company this site is, shows the same split: the strongest sites tend to carry it, the weakest do not. Nothing in the rubric penalises being big as such; a ten-page site can score in the 80s. What it penalises is entity incoherence, and the large estates in our scan rarely held coherence, because scale is what funds the migrations, handovers, acquisitions and multi-brand structures that fracture one company into many.

Below about fifty-five, the engines we prompted began to omit or invent

Here the distinction gets practical, and it needs stating carefully. A general-purpose model already knows from its training data that a household FTSE 100 name sells oil or makes medicines; you do not need clean markup for that, and this piece does not claim otherwise. The exposure is narrower and sharper. It surfaces in the engines that read your live site to answer, ChatGPT with search, Perplexity and Google's AI Overviews, and in the specifics that were never reliably in the training data anyway: your current product range, the sub-brands, the divisions, and exactly who each one is for. That is the layer the AI Search Readiness score measures. It reads each site the way those retrieval engines do, resolving the company as one entity, extracting what it sells and to whom, and checking whether the facts survive being repeated. We set the bands against what the engines actually returned when prompted on each site. Above 70, they resolved identity and quoted products cleanly. Between 55 and 69, identity held but product extraction was patchy. Below 55, they routinely omitted products or stated them wrongly, and on the weaker sites invented a buyer that did not exist. Fifty-five is not a law of nature; it is where the answers stopped being trustworthy, and it is the commercial line, because a buyer asking an engine to shortlist suppliers wants your offer, not your name. In this scan, 38 of the 40 readable sites sat below 55 on the Crane Index.

Here is a test you can run in a minute, and it is deliberately falsifiable. In Perplexity, or ChatGPT with search turned on, paste this: "Using only [yourdomain.com], name this company's three main products or services and the specific buyer each one is for." Forcing it to the live site is the point, because a model reciting old training data is not the test. On a site above 70 it will answer cleanly. On a site below 55 it will typically miss or muddle at least one, and on the weakest it will invent a buyer that does not exist. Run it on yourself, then on a competitor. Much of what moves the score is whether the site carries valid Schema.org structured data, chiefly the Organization and Product types as JSON-LD, but it is not only a markup check: it also weighs whether one name and identity resolve consistently across the estate, which is what lets an engine see a single company rather than several.

The field clusters low, and what separates the top is structure, not spend

Across the 40 readable sites, the Crane Index ran from 21 to 58. Not one site reached 70, 2 sat between 55 and 69, 22 between 40 and 54, and 16 below 40.

The mass sits in the middle and lower bands, so this is not a ready field dragged down by a few neglected outliers. It is how large corporate sites are built and signed off: comfortable for humans, illegible to machines. And the search read was not the weakest. The Crane Index takes three reads of every site, and the other two frame it. Agent Readiness, whether an automated agent can reach and act on the site at all, was the strongest at a median of 64. Brand Representation, whether a machine can confirm from your own site who you are, what you do and which real-world entity you correspond to, was the weakest by a distance: a median of 15 out of 100. The country's biggest companies are, on their own websites, largely unable to prove to a machine who they are.

The full table: every site a machine could read, scored and ranked

Here is the field, named. Rank is by the Crane Index, the composite of the three reads; the three scanner columns follow. These are measurements of each company's primary corporate site as an automated reader found it on 6 August 2026, nothing more: a low score says the site is hard for a machine to read, not that the business is anything other than excellent at what it does. The 39 sites our reader could not get into at all are excluded as unmeasured rather than scored zero, which is a courtesy the averages above extend to every company here. How to read this table, and the full method →

Read the columns against each other and the story sharpens. Agent access is broadly survivable; representation is where nearly everyone collapses; and the composite is dragged to a median of 42 by the read that says who you are. Not one of the country's hundred largest companies clears 70 on the Crane Index.

The good news hides in the same data. The low-band sites hold the same commercial facts as the high-band ones, so remediation is cheap. You are not writing new content, you are making existing content legible. The habit that separated the top of this field, valid Organization markup and a consistently resolvable identity, is naming consistency and accurate structured data, not a rebuild, which means the median can move faster than a board expects.

The sites we could not score are a finding, not a footnote

The fair challenge is that we scored 40 and set aside the rest, which sounds like defining away the hard cases. Walk through who was set aside. Of the 100 constituents:

• 21 we could not resolve to a single scoreable parent domain (holding structures and multi-brand groups chief among them), and a company an engine cannot resolve to one domain is already an AI-legibility failure.
• 23 answered an automated reader with an outright refusal.
• 9 gave it nothing at all.
• 7 serve pages that are empty until scripts run, which is how a site looks to the many machine readers that do not execute them.

The constituent list we used is pinned alongside the archived results, so the exclusions are inspectable rather than convenient. None of those postures is neutral: the fetchers that answer engines send when a user asks about you live in the same data centres our reader does. One honest caveat cuts the other way: our reader is a single unfamiliar bot, and a site that verifies crawlers by network may admit the engines it recognises while refusing ours, so treat the blocked 39 as unmeasured, not as failed. That is exactly why they are excluded from every median above. The 47 median is drawn only from the sites that let a machine in, the most legible end of the index by construction. The full picture is unlikely to be better than what we report.

Three honest limits sit beside that. Each read is of the homepage and the site's root signals, not a crawl of every inner URL, so a large estate may carry cleaner markup on a product page our reader did not open, and this is a read of the front door. The 40 scored sites are a sample, not a census. And a single scan is a snapshot of engines and estates that shift month to month, so read the bands as a strong signal rather than a fixed constant. The full result set, every company, every check, is archived and dated, and we will run the same scan again and publish the movement.

What to do now: put one legibility gate in the deploy pipeline

Stop inferring health from traffic and put a hard check in front of release. First, block any deploy where the Organization schema is missing, invalid, or renders a different legal name, URL or identifier across templates. That is the single habit the data separates cleanest on: present and valid on the strongest sites, absent on the weakest. Second, require valid Product markup, matched to the visible copy, on every page that carries commercial weight, so an engine can quote the offer rather than guess it. Third, give one senior owner the score and the mandate, because large sites drift into the lower bands the moment legibility has no owner and every migration erodes it further. This is AI search discipline: entity clarity and structured data decide whether you are read, trusted and named. Most of the lift comes from accurate Organization and Product schema in JSON-LD, not exotic markup.

The stakes are compounding

Here is the piece of this a board should feel in the P&L. When a buyer asks an engine to shortlist suppliers, the model does not weigh your media budget, your brand tracker or your analyst coverage; it weighs whether your site states, in machine-readable form, what you sell and to whom. If it cannot, the shortlist slot goes to a legible competitor at no cost to them, and your paid channels are left buying back demand you should have received free, at a cost per acquisition that only rises. Illegibility is not a technical debt line; it is a standing subsidy to whoever in your category fixed their markup first.

The shift is no longer a forecast; we are inside it. In 2026, Google reported that its AI Mode in Search had passed a billion monthly users, and the scan above is what that shift found waiting. When the engine names a shortlist, whether you make it turns on how cleanly you are written for a machine, in the structured data Google documents, not on how large you are. The web rewarded whoever could publish the most pages; the answer engines reward whoever a machine can resolve as a single company and quote without error, and on this cohort those are seldom the same names. The same instrument found the same mechanism in retail: see the UK's biggest online retailers benchmark, and the full programme at Crane Index Research.