# The Biggest Names in Distribution Score Worst on AI Readability

> We graded 984 industrial catalogues, then re-measured 25 household names live. Their median is 40 against a market median of 50. The full named table.

**Published:** 2026-08-13  
**Category:** Research · **Tags:** AI visibility, benchmark, industrial distribution, robots.txt, structured data, GEO  
**Canonical:** https://partsgraph.ai/blog/biggest-distributors-worst-ai-readability

## In short

**How do the largest industrial distributors and manufacturers score on whether AI assistants can read their product catalogues?**

Worse than the market they lead. Across 984 graded industrial catalogues the median score is 50 out of 100. Re-measuring 25 household names live on 13 August 2026 gives a median of 40, with 15 of the 25 falling below the cohort median. RS Components scores 35, Farnell 33, Mouser 34, Grainger 34, ABB 34 and Zoro 31, while the highest score anywhere in the full cohort — 80 — belongs to a small specialist distributor nobody outside its niche has heard of. Scale appears to be a liability here rather than an advantage, because large estates accumulate the exact things that block machine access: aggressive bot rules at the edge, JavaScript-rendered specifications, and catalogue pages that no crawler is permitted to reach.

---

The Industrial AI Visibility Benchmark measured 984 manufacturer and distributor catalogues on a single question: can an AI assistant actually retrieve a product specification from this site? The median score was 50 out of 100.

That number was published anonymised. This article is not anonymised, because the more useful finding was hiding inside the distribution.

**The largest, best-known companies in industrial distribution score below the median of the market they lead.** Their median is 40. Fifteen of the twenty-five household names we re-measured fall under the cohort's own midpoint, and the highest score anywhere in the full 984 — an 80 — belongs to a small specialist distributor with no brand recognition outside its niche.

## The table

Measured live on **13 August 2026**, not lifted from the earlier fieldwork. Every figure below can be reproduced by anyone using [the free grader](https://partsgraph.ai/audit), including by the companies named.

| Company | Domain | Score | Grade |
|---|---|---|---|
| Voltimum UK | voltimum.co.uk | 78 | B |
| Rexel UK | rexel.co.uk | 66 | C |
| Siemens | siemens.com | 65 | C |
| Cromwell | cromwell.co.uk | 65 | C |
| B&Q | diy.com | 65 | C |
| Toolstation | toolstation.com | 60 | C |
| Jewson | jewson.co.uk | 62 | C |
| Reichelt | reichelt.de | 61 | C |
| Legrand | legrand.com | 58 | C |
| Screwfix Trade | screwfix.com | 50 | D |
| Digi-Key | digikey.com | 49 | D |
| HARTING | harting.com | 42 | D |
| TE Connectivity | te.com | 40 | D |
| Schneider Electric | se.com | 39 | F |
| Festo | festo.com | 39 | F |
| RS Components UK | uk.rs-online.com | 35 | F |
| Würth | wurth.com | 34 | F |
| Travis Perkins | travisperkins.co.uk | 34 | F |
| Mouser | mouser.com | 34 | F |
| ABB | abb.com | 34 | F |
| Grainger | grainger.com | 34 | F |
| Sonepar | sonepar.com | 34 | F |
| Conrad | conrad.com | 34 | F |
| Farnell UK | uk.farnell.com | 33 | F |
| Zoro US | zoro.com | 31 | F |

Median of these 25: **40**. Median of the full 984: **50**. Range here: 31 to 78.

## Why we re-measured before publishing

The benchmark fieldwork ran on 10 August. Publishing a named score three days old would have been careless, so every company in the table was re-graded on the morning of publication.

**Twenty-one of the twenty-five reproduced exactly.** That is a reassuring result for the instrument. Three did not:

- **Rexel UK: 57 → 66.** A nine-point improvement in three days.
- **TE Connectivity: 53 → 40.** A thirteen-point fall.
- **Schneider Electric: 51 → 39.** A twelve-point fall.

Nobody at those companies is likely to have been working on AI readability that week. A change to an edge rule, a WAF policy update, a CDN configuration pushed for an unrelated reason — any of these can move the score double digits overnight, in either direction, without anyone noticing.

Then it happened again while this article was being written. A second pass a few hours after the first found **Toolstation had moved from 63 to 60** — a smaller change, but inside the same day, and the table above carries the corrected figure rather than the one drafted that morning.

That is the most operationally useful thing in this article. **Your machine readability is not a property you fix once.** It is a live configuration that drifts, and it drifts silently, because nothing in your monitoring watches for it. Four of these twenty-five moved within seventy-two hours, one of them within a single afternoon, and it is safe to assume nobody at those companies knew.

## What actually separates 78 from 31

The four things measured are not equally hard to fix, and the gap between top and bottom is rarely about the top scorer having done something clever.

### Crawler policy — the cheapest points on the board

Most sites in the cohort have a robots.txt that says nothing whatsoever about AI crawlers. It is not hostile; it is silent, written before this category of client existed. Silence increasingly resolves to refusal, because the safe default for a crawler facing an unstated policy is to stay out.

Naming the answer-engine crawlers you welcome is an afternoon's work and it is the single highest-leverage change available to almost every company in the table.

### Edge reachability — where the big estates lose

This is where scale actively hurts. Enterprise bot management exists to stop scraping, credential stuffing and price harvesting, and it is good at it. It is also, by default, suspicious of any client that does not look like a browser — which describes every answer-engine crawler.

Several of the lowest scores in that table are not content problems at all. The content is fine. A non-browser client simply never reaches it.

### On-page readability — the JavaScript trap

If your specifications are assembled in the browser by a JavaScript application, most crawlers see an empty container. The product data exists, is accurate, and is invisible. Server-rendered structured data — Schema.org `Product` with real attributes, not a name and a price — is what closes this.

### Agent surfaces — where essentially nobody has started

Across all 984 catalogues, **not one exposed an endpoint an agent could query directly.** No MCP server, no machine-readable feed designed for retrieval rather than for a shopping comparison engine. The bar here is not high. It is simply that almost nobody has cleared it, which makes it the cheapest available differentiation for anyone who moves first.

## What this is not

It is worth stating the limits of the claim, because a benchmark that oversells itself deserves the scepticism it gets.

This is **not** a measure of catalogue quality. Several companies with poor scores have excellent product data. The problem is access, not accuracy.

It is **not** a claim that a low-scoring brand is absent from AI answers. Large brands are described extensively across the web, and an assistant will happily talk about them using third-party sources. The risk is more specific: the assistant answers a technical question about your product using a distributor's data, an old cached page, or a competitor's comparison table, and cites them instead of you.

It is **not** a fixed ranking. As three companies demonstrated within three days, these scores move.

And it is **not** an accusation. Nothing in the table is the result of a decision anyone made badly. It is the accumulated residue of sensible individual choices — block the scrapers, render in the client, ship the SPA — taken before a new kind of reader arrived.

## If you are in the table

Run [the grader](https://partsgraph.ai/audit) on your own domain. The breakdown shows which of the four probes failed and precisely why, and it is free with no signup because an argument you can check yourself is worth more than one you have to take on trust.

If you think the result is wrong — a regional storefront misidentified, a geo-redirect followed incorrectly, a rate limit tripped — tell us. We will re-run it and correct the record publicly. We have done that before. A measurement nobody is willing to be wrong about in public is not a measurement, it is marketing.

## Method

Each site is probed as an AI crawler would approach it. We fetch the homepage and robots.txt, resolve the catalogue through the sitemap, and sample product pages with a standard non-browser client. Four dimensions are scored: crawler access policy, machine reachability at the edge, on-page machine readability, and agent surfaces. Nothing behind a login is touched, no rate limits are deliberately exceeded, and a site that blocks us is recorded as blocked rather than retried in disguise — the block is the finding.

The full cohort of 984 remains published anonymised as [the Industrial AI Visibility Benchmark 2026](https://partsgraph.ai/research/ai-visibility-benchmark-2026). The 25 companies named here are named because they are large, publicly traded or widely known, the measurement is reproducible by anyone, and the finding is materially more useful with the names attached than without them.

## Common questions

### How were these scores measured?

Each site is probed the way an AI crawler would approach it: we read the homepage, robots.txt and a sample of catalogue pages using a standard non-browser client, then score four things — whether the crawler policy names AI agents, whether non-browser clients get through the edge at all, whether product specifications are in server-rendered structured data or trapped in JavaScript, and whether the catalogue is discoverable and reachable. No login, no crawling behind authentication, nothing a search engine would not also attempt. The same grader is free and public, so any figure here can be reproduced by the company it describes.

### Is a low score the same as being invisible to ChatGPT?

No, and it is worth being precise. A low score means an assistant has to work harder to retrieve accurate specifications from your site, and is likelier to fall back on a distributor, a marketplace listing or an older cached description. It does not mean your brand never appears. The practical risk is subtler than invisibility: the assistant answers a specification question about your product using somebody else's data, and cites them.

### Why would large companies do worse than small ones?

Three compounding reasons, all of them consequences of scale rather than carelessness. Large estates put a bot-management layer at the edge, and the default posture of those products is to refuse unfamiliar clients — which now includes answer-engine crawlers. Large catalogues are usually rendered by a JavaScript application, and most crawlers do not execute JavaScript. And large sites accumulate robots.txt rules over years, written to stop scrapers and price bots, that were never revisited when a new category of legitimate crawler arrived. None of that is a mistake anyone made on purpose. It is drift.

### The scores in your August benchmark differ from these. Why?

Because we re-measured on the day of publication rather than reusing fieldwork. Of the 25 companies named here, 21 reproduced their earlier score exactly. Three moved by more than three points in three days: Rexel UK rose from 57 to 66, TE Connectivity fell from 53 to 40, Schneider Electric fell from 51 to 39. That volatility is itself a finding — a single edge-configuration change can move a score double digits overnight, which is why a one-off audit has a short shelf life.

### Our score looks wrong. What do we do?

Run the grader yourself and read the breakdown, which shows exactly which probe failed and why. If you believe the result is a false negative — a regional storefront misidentified, a geo-redirect we followed incorrectly, a rate limit we tripped — tell us and we will re-run it and correct the record. We have made that correction before and will again. The measurement is only worth publishing if we are willing to be wrong about it in public.

### Is this just SEO with a new name?

It overlaps but it is not the same problem. Classical SEO optimises for a ranked list of links a human clicks. What this measures is whether a machine can retrieve a specific technical fact — a rated current, an ingress rating, a thread pitch — and reproduce it accurately enough to cite you. A site can rank well and still be unreadable in that sense, which is exactly what several companies in the table below demonstrate.

## Sources

1. [The Industrial AI Visibility Benchmark 2026 — 984 graded catalogues](https://partsgraph.ai/research/ai-visibility-benchmark-2026)
2. [Partsgraph grader — free, no signup, reproduces any figure in this article](https://partsgraph.ai/audit)

---

Partsgraph — the agent-ready parts data layer. Free AI-visibility grader: https://partsgraph.ai/audit