Thinking · 18 September 2026

The Brochure Tax

The cost of a site AI cannot read is real, ongoing, and almost entirely invisible, because the things it costs you never appear in your analytics.

The Brochure Tax is the ongoing, invisible cost of a static brochure site that AI cannot read or act on: the referrals never made, the questions never answered, and the enquiries handled by hand that should not have needed a hand.

It is called a tax because it is not a one-off. It is charged every month, it rises quietly as more buyers route their decisions through assistants, and nobody sends you an invoice for it.

Why it is invisible

Every other marketing cost announces itself. An ad spend appears on a card statement. A bad month shows up in the numbers. The Brochure Tax is different, because what it costs you is things that did not happen, and things that did not happen leave no trace.

When someone asks an assistant to recommend a supplier and your business is not among the three it names, there is no impression, no click, no bounce, no session. Your analytics will show a perfectly ordinary week. The buyer will contact somebody else, and the only honest description of what you can see is: nothing at all.

That is the whole difficulty with this problem. It is not that businesses ignore it. It is that the reporting they trust is structurally incapable of showing it to them. Every instrument on the dashboard measures people who arrived. None of them measures people who were told to go elsewhere.

The three components

Referrals never made

The largest line, and the one you can see least. If a model cannot resolve what you do, who you serve and on what terms, it will not name you when asked. It is not penalising you. It is answering with whatever it could actually read.

There is a mechanical version of this failure that has nothing to do with content quality. An analysis by Vercel and MERJ of a month of crawler traffic (569 million GPTBot requests and 370 million from ClaudeBot) found that none of the major AI crawlers execute JavaScript. If your site renders its content in the browser, the crawler receives an empty shell. Your copy may be excellent. It is simply not there when the machine looks. A business can be paying the maximum possible Brochure Tax while having a site everyone internally agrees is rather good.

Questions never answered

A visitor with a question at eleven o’clock on a Sunday either finds the answer on the page or leaves. Most brochure sites defer the answer: pricing says “every project is unique”, delivery says “get in touch”, scope says “bespoke”. Each deferral is a small toll, charged to a visitor who had already decided to be interested.

The deferral is rarely accidental. It is usually a considered decision to move the conversation to a call, where it can be handled properly. That works when the buyer is willing to have the call. It fails silently when they are not, and you never learn which of the two happened.

Work done by hand that should not be

The component you can actually measure. Every enquiry that exists only to ask something your site could have answered, every call taken to repeat your opening hours, every quote email that restates a price you could have published: that is your own time, billed at your own rate, spent on work a page could have done once.

How to audit your own, in thirty minutes

This is the part worth doing, because it turns an argument into a number that belongs to you rather than to an article.

1. Pull your last twenty enquiries Email, form submissions, whatever you have. Twenty is enough; you are looking for a pattern, not a p-value.
2. Mark each one: could a better page have answered this? Be strict. “How much does X cost” and “do you cover my area” are yes. “Here is my situation, what would you do” is no: that is the enquiry you actually want.
3. Estimate the handling time for the yes pile Including the reading, the reply, and the second reply after they come back. Most people undercount this by half.
4. Multiply by your own hourly cost, then by your enquiry volume That is the measurable component, per month, in pounds. It is usually not enormous. It is also usually enough to pay for fixing the pages that caused it.
5. Now check the two you cannot count Open your own site’s raw source and see whether the content is there without JavaScript. Then ask three assistants who they would recommend for what you sell, and see whether you are named.

The last step is not measurement and should not be presented as such. It is a spot check, and the sample is small and the answers vary between runs. It is still more information than you had, and if you are absent from all three answers on a question you should obviously feature in, that is worth knowing.

A worked example

The figures below are illustrative, not a client. They exist to show the shape of the calculation, and the only number you should trust is the one you produce from your own twenty enquiries.

Take a firm doing forty enquiries a month. Of the last twenty, nine were asking something the site could have answered: what it costs, whether they cover the area, how quickly they can start. That is 45%, so call it eighteen a month. Each takes twelve minutes across the first reply and the inevitable second, which is three and a half hours a month of somebody senior answering questions a page could have answered once.

That is the visible part, and on its own it is a middling annoyance rather than a crisis. The point of the exercise is what it implies about the invisible part. If nearly half the people who did make contact had to ask a question the site should have answered, it is reasonable to assume some number of people who did not make contact had the same question and simply left. You cannot count them. You can infer that they exist, and their behaviour is not going to be better than the behaviour of the people who bothered to email.

Then the third layer: whatever an assistant said about that firm when someone asked. If the pricing is not on the site, the assistant cannot state it, and a buyer comparing three suppliers gets two with numbers and one without.

Why it compounds

Two reasons, and neither is dramatic.

First, the share of decisions passing through an assistant is rising rather than falling, so the same unreadable site costs more each year than it did the year before. The rate goes up while the site stays still.

Second, citation is self-reinforcing. Once engines have learned to treat a particular source as the one that answers a category’s questions, that source is the easiest one to reach for next time. The gap between being cited and not being cited widens on its own, without anyone at either business doing anything. The competitor who fixed this in 2025 is not merely ahead; they are accumulating an advantage that gets cheaper for them to hold.

There is a third, duller mechanism worth knowing about. The same Vercel and MERJ analysis found that ChatGPT’s crawler spent 34.82% of its fetches on URLs returning 404, and Claude’s 34.16%, against 8.22% for Googlebot. Roughly a third of the attention these crawlers bring to the web lands on pages that no longer exist. On a site with years of accumulated dead links and moved pages, a meaningful share of whatever attention you get is spent discovering that your old URLs are gone.

What it is honestly worth

A note on rigour, because this is where this sort of argument usually goes wrong. You will see confident figures attached to problems like this: a percentage of revenue, a cost per missed referral. Treat them with suspicion, including any we might be tempted to offer. The referral component is genuinely unmeasurable from where you stand: you cannot count conversations you were never part of.

What you can measure is the third component, and you can reason about the second from your own bounce behaviour. That is enough to decide. It does not need a fabricated total to be worth fixing, and a number invented to make the case would undermine the case the moment anyone checked it.

When you are not paying it

This does not apply to everyone, and it is worth saying so plainly rather than implying that every business is bleeding money.

If you sell entirely through relationships and referral, if your work comes from a framework agreement or a handful of long-standing accounts, or if your buyers find you through a channel that has nothing to do with search (a marketplace, a trade body, a directory they already trust), then your website is a credential rather than a channel, and the tax on it is small. Make it accurate, keep it fast, and spend the money somewhere it does more.

The tax is large when a meaningful share of your buyers start by asking a question of someone or something, rather than by already knowing your name.

How to stop paying it

The fix is structural and mostly unglamorous: make sure the content is in the HTML, resolve the entity, publish the answers, give the machines a source of truth, and put something on the site that can answer at eleven on a Sunday. That is the definition of an AI-native website, and it is substantially cheaper than the rebuild people assume it implies, because three of its four properties can be applied to the site you already have.

Ask what yours is costing Run the answer-ready test

Sources

• Vercel and MERJ, The rise of the AI crawler: one month of crawler traffic across Vercel’s network, including 569 million GPTBot requests and 370 million ClaudeBot requests.
• Jeremy Howard, llmstxt.org: the llms.txt proposal, first published September 2024, second version August 2026.