The assistant answering your customer is reading your website.

What did OpenAI actually say?

On 8 September 2026 OpenAI published a business update called The Work Now Within Reach. There is nothing in it to download and no feature to switch on, which is exactly why it is worth reading. It is the company explaining what its newest model is good at and who pays for the version your customers use.

The sentence that matters if you own a website is short. OpenAI says GPT-6 Astra is state of the art in areas such as computer use, browsing, software engineering, cybersecurity, science and professional work. Browsing is on that list. Browsing is the bit that opens your homepage while somebody waits.

Two other lines deserve reading twice. OpenAI says its products reach more than one billion weekly active users and 2.5 million businesses. And on the money it is unusually blunt: free access, supported by advertising, helps people discover where AI is useful. Those three sentences, taken together, describe a machine that reads live web pages, at enormous scale, for free.

Why does browsing matter more than the model being clever?

There is a difference between a model that remembers your business and a model that reads it. A model answering from memory works with whatever was in its training data, which for a hairdresser in Faversham or a groundworker in Hastings is usually nothing at all, or a directory entry from four years ago with the old phone number on it.

A model that browses goes and looks. Now. At the pages it can reach in a couple of seconds.

That flips which of your problems is the expensive one. Being unknown stops being the main risk, because the model is not relying on having heard of you. Being unreadable becomes the main risk. A homepage that says "quality service since 2009" and never says what the service is. A prices page that is a PDF. A phone number drawn into a header image. An opening hours block that has said "Christmas hours coming soon" since February.

When the model browses, whatever is on the page is the answer. It will not ring you to check, and it will not give you the benefit of the doubt. If your page does not say you cover Tonbridge, then as far as the customer is concerned, you do not cover Tonbridge.

Who is asking, and does the asking stop?

The scale line is easy to wave away, so look at the behaviour instead. In OpenAI's own study of people on individual ChatGPT plans, daily message volume was roughly 50% higher six months after signup than in the first month, and people had tried roughly twice as many distinct tasks. Keep the hedges attached: roughly, and in a study of its own users. Read it as direction rather than gospel.

Direction is enough to act on. It says people do not settle into one use and stop. They start with something safe, an email they cannot face writing or a recipe for what is in the fridge, and six months later they are asking it to find a vet that will see a cat on a Sunday, or whether the garage on the high street does MOTs on a Saturday morning.

That widening is the whole story for a local business. The habit grows into exactly the practical, boring, local questions that used to go into a search box. It is the same slow absorption we described in how ChatGPT recommends local businesses: not a replacement for Google in one go, just one category of question at a time, quietly.

What does supported by advertising change for local search?

OpenAI states it without decoration: free access, supported by advertising, helps people discover where AI is useful. The free tier is the one your customers are on. It is now a funded product rather than a loss leader waiting to be withdrawn or throttled.

Be careful what you take from that. OpenAI has not said adverts will appear beside local recommendations, and we are not going to pretend it has. What it has said is that the free, browsing, one-answer experience has a business model underneath it, which means it gets more usage rather than less, and that its answers land in front of people who are not paying, are not comparing ten blue links, and will mostly act on the first name they are given.

The practical version: the cheapest and most-used surface in AI search is the one that reads your live website and hands a customer a single answer. You cannot buy your way into it today. You can be legible to it today, which is a better deal than it sounds.

Why can a free answer afford to read your whole site?

Because reading got cheaper. OpenAI says GPT-5.6 Sol helped improve its production serving software, reducing end-to-end serving costs by 20%, with further improvements raising token-generation efficiency by more than 15 percent. That is a company reporting on its own infrastructure rather than a benchmark anyone outside has checked, so treat the precise figures as its own accounting.

The direction still matters to you. When an answer costs less to produce, it can afford to be longer and to open more pages before it commits to a recommendation. A model that can only afford one fetch takes whatever your homepage happens to say. A model that can afford five will find your services page, your area page and your prices, and the business with those pages wins the comparison it was never told it was in.

The same update gives a sense of how much work is being handed to machines rather than people. OpenAI says its research organisation now uses 3.1 agent-workdays of effort for every workday of human labour. That is inside OpenAI, doing research, and it is not a claim about anybody else's business. It is still a fair signal of where the effort is going, and of how casually a machine will now be sent off to read twenty websites on somebody's behalf.

What does a browsing model need to find on your page?

Nothing clever, which is the annoying and reassuring part. The list has barely changed in two years.

• What you do, in plain words, high on the page. "Emergency plumber and boiler repairs" beats "your local solution partner", because one of those can be quoted in an answer and the other cannot.
• Where you do it, named. Towns and postcode areas written out as text, not implied by a map embed or a radius nobody can see.
• A price, a range or a call-out fee. A page with no number on it cannot appear in an answer about cost, and cost is most of what people ask.
• Hours that are true this week, bank holidays included, as text rather than as a graphic.
• A phone number and an address in the HTML, not baked into a header image or loaded by a script after the page settles.
• One real page per service and per town you genuinely serve. Thin duplicates with the town name swapped are worse than not having them, because now the model has read them.

If that looks like the checklist from our post on Google's AI phoning local businesses, it is. Different assistant, same homework, which is the best news in this entire article.

What are we doing for the sites we look after?

Unglamorous things, on purpose.

• Reading each client homepage the way a machine would, with JavaScript switched off, to see what actually survives.
• Getting the service and the area into the first screen of copy as text, rather than leaving them in a slider or a heading image.
• Putting a starting price, an hourly rate or a call-out fee on the main service page for every client who will agree to one.
• Checking that phone number, address and opening hours sit in the page HTML and match the Google Business Profile exactly, because a model that finds two versions will pick one and it may not be yours.
• Cutting thin town pages instead of adding more of them. A browsing model reads one page properly, and a shallow one now embarrasses you in public rather than just sitting there.

None of that is an AI project, and we would have recommended most of it in 2019. What has changed is the reader: less forgiving, faster, and speaking directly to your customer. You can see what that work turns into on our case studies.

What should you do this week?

1. Open your own homepage on a phone, on mobile data, and read only what appears before you scroll. Write down the answer it gives to "what do they do and where". If you cannot write it down, a model cannot either.
2. Ask ChatGPT, with browsing on, what your business does and who it serves. Judge the answer against your page, not against your opinion of your business.
3. Put one honest number on your main service page: a starting price, an hourly rate or a call-out fee, with the thing it depends on.
4. Write your towns out as words on the page. Not a map, not "we cover the South East", the actual names of the places you will drive to.
5. Check your opening hours in three places: your website, your Google Business Profile and your front door. Fix whichever one is lying.
6. Pick the thinnest page on your site and either make it genuinely useful or delete it before something reads it aloud to a customer.

Three objections, answered

"This is all enterprise stuff. None of it is about a barber in Faversham." Largely true, and worth saying out loud. The customer stories OpenAI leads with are big organisations, starting with Boston Children's Hospital in the United States, where AI-assisted research is credited with more than 40 diagnoses in previously unresolved rare disease cases. That is not a local business story and we will not dress it up as one. The read-across is not the case study, it is the capability. The same browsing model that reads a research paper reads your prices page, and the version doing that is free.

"Every year somebody says the model is smarter. It was noise last time." Much of it was. The specific claim on the table is bigger than usual: OpenAI says one of its internal models produced a solution to the Navier-Stokes Millennium Prize Problem, open for roughly 90 years. Take that as a company describing its own result, hedge included. You do not have to believe any of it to act on the browsing part, because you can test that yourself in a minute with your own website and a free account.

"If the AI answers the question, nobody reaches my website, so why improve it?" This is the objection that argues against us and it deserves a straight answer. Some visits do disappear, and the sessions graph in your analytics is not going to flatter you. What replaces them is being named, with your number, in an answer given to somebody who has already decided. You cannot be named out of a page the model could not read, so the work is identical either way. The traffic argument only decides how cheerful you feel while doing it.

The verdict

Strip the announcement back and a local business is left with two facts it can act on. OpenAI says its newest model is best in class at browsing the live web, and that free access is funded by advertising, so the free tier is not a trial that quietly ends in the spring. It has not said which model free users get. Our read is that it matters less than it sounds, because browsing is how these assistants answer now.

That makes your website a live document, read by a machine on a customer's behalf, probably several times a week, without anybody telling you it happened. The businesses that come out of the next year well will not be the ones with an AI strategy. They will be the ones whose pages say what they do, where they do it, when they are open and what it costs, in text, on the page, today.