Somebody at Commonwealth Bank has written a strategy into a robots.txt file.

It sits under a heading about AI and LLM crawlers, and it names seventeen of them, GPTBot and ClaudeBot and PerplexityBot and the rest, and it lets all of them in. Above the list is a comment explaining why. If a model is going to answer a question about the best home loan in Australia, "we want it answering from us."

The file was last reviewed on 4 August 2026. Somebody maintains it. In an index where most of the two hundred are running whatever file arrived with the hosting, that comment is the most deliberate sentence about machine readership anyone on the ASX has published.

It is also in the one place nobody in the C-suite will ever look.



There are two versions of your company

The first is the one you write. Annual report, results presentation, the call, the site. You control every word of it, you spend real money on it, and you have a team whose job is to get it right.

The second is the one you emit. Job ads. Patent filings. Headcount. Capex lines. Signed customers with dates attached. Nobody in your organisation thinks of this as communication, because it isn't. It is exhaust.

Both versions are being read. They are not being read by the same people, and they do not say the same thing.



The market cannot tell them apart for about a year

There is now a decent measurement of this. A 2025 working paper out of Florida separates what it calls AI talk from AI walk across 721 US listed companies and twenty thousand firm-quarters, from 2016 to the middle of 2024. Talk is forward-looking AI claims pulled out of earnings call transcripts. Walk is actual AI expertise, read off employee resumes.

Only walk predicts anything. It predicts patent volume and patent quality. Talk predicts neither, and within a given company, past talk does not forecast future walk. The two are unrelated.

The market pays for talk anyway. Short-run abnormal returns go to the companies that say the most. By three hundred and sixty days the ranking has inverted, and the talkers underperform while the builders take a persistent premium.

Institutional investors sort the two early. The broad market does not sort them at all.

None of this is new, it is just better instrumented. In 1998 and 1999, ninety-five US companies changed their name to a dotcom. They earned cumulative abnormal returns of fifty-three per cent over the five days around the announcement. The effect was identical whether or not the company had any internet business, and the ones with no internet business at all went on to earn two hundred and forty-three per cent over the following four months.

A label, requiring no operational change whatsoever, produced a permanent revaluation. The machine reading it did not care what was behind it, because the machine could not see what was behind it.



So I went to build the measure

The obvious next move is to score both sides and publish the gap. Talk score, walk score, signed distance between them, ranked within sector. We have the talk side already, across the whole ASX 200.

The walk side does not want to be read.

Google Patents refuses crawlers at robots.txt. The PatentsView API has migrated and now needs a key. IP Australia's patent API needs registration, and its bulk data is a paid request. Seek, which holds the richest hiring dataset in the country, names seven AI crawlers individually and blocks them from two paths: the job listings, and the company pages. Corporate career pages render their listings in JavaScript, so there is nothing on the page for a machine to read. Headcount is a number inside a PDF.

Every one of those is a deliberate decision by somebody. Together they add up to a market where the evidence is gated and the narrative is free.



Which is the opposite of what everyone assumes

The comfortable version of this argument says the serious money reads the hard data, so your communications are decoration. Stop worrying about the website.

The gating says otherwise. The hard data costs money, needs credentials, and in several cases is actively defended. That is precisely why it confers an advantage, and precisely why only the firms who can pay for it are using it.

Everyone else, which is most of the market and nearly all of the retrieval layer now sitting between the market and your filings, reads the free half. Your half. The one you left on a corporate site that half the index has never checked is reachable.

Your narrative surfaces are not the slow channel. For most of the money, they are the only channel.

And visibility is priced. Merton established the model in 1987: investors will not hold what they do not know exists, so a narrower investor base carries a higher required return. Grullon, Kanatas and Weston put numbers on it in 2004, across nearly six thousand firm-years, using advertising spend as the proxy for how widely a company is known. A one standard deviation increase associated with ninety-nine per cent more shareholders, twelve per cent more institutions, spreads four and a half per cent narrower, quoted depth twenty-five per cent deeper, and price impact eighteen per cent lower.

This is a brand budget arriving as a cost of capital line. Advertising was the measurable proxy available in 2004, and part of that effect is simply that consumer brands collect retail holders. The mechanism underneath it is not advertising. It is being findable by someone deciding where to put money. That used to mean a television buy. It now also means whether a machine can open your results.



What the ASX has actually done about it

Eighteen of the two hundred restrict machine access. One, Steadfast, closes the door completely. Three block the investor material specifically: TPG Telecom disallows the directory holding its ASX announcements and every PDF on the site, Chorus disallows all PDFs, and Greatland Resources names three files to keep out, two of which are the prospectus and the corporate presentation.

Five go the other way and name AI crawlers in order to let them in. Commonwealth Bank names seventeen, NextDC eighteen, Cochlear sixteen, DroneShield more than twenty.

The rest have a default file they inherited with the hosting.

The clearest illustration in the index is Flight Centre. Its consumer site carries a considered policy, written by somebody who had thought about it, allowing GPTBot and AnthropicBot and PerplexityBot and Applebot by name. Its corporate and investor site serves an empty file.

Marketing decided how machines should read the holiday packages. Nobody decided how machines should read the results.



The trap

You are being scored on both versions of yourself, on different clocks, by different money.

The version you author is free to fix and nearly free to publish, and most of the index has never checked whether a machine can reach it. The version you emit is expensive to read, which is the only reason it still confers an edge.

If the two say different things, the gap is not a communications problem. It's a twelve-month clock: talk gets the short-run price, walk gets the persistent premium. The people who can afford the walk data already know which way it runs.

Plan B reads the machine-legible surface of listed companies and reports the variance against the story they tell. The robots.txt findings above come from a read of all two hundred ASX 200 constituents, September 2026.