Here’s the question I get more than any other right now: if AI assistants and agents do the searching and the buying, do businesses still need a website? My answer is yes, more than ever, but not every business, and not the kind of website most businesses have. Every brand needs a home it owns and a place to publish the truth about itself. You can’t control what ChatGPT or Gemini says about you. You can control what your website says.
The short answer
- Your website is the one place you control the facts about your business. AI answers are summaries of other people’s pages, and they get things wrong.
- AI gets brand facts wrong often. In a Seer Interactive test, AI models gave a phone number for 91% of branded questions, and 36% of those numbers didn’t match the brand’s own page.
- The AI companies publish their own product record. Their official websites and changelogs document capabilities, limits and releases. Buyers and assistants can check that record.
- Not every business needs one. About 55% of employing businesses in the ABS survey reported their own website or app in 2024-25. The survey has scope exclusions and does not describe every Australian sole trader.
- The website becomes a premium asset, built for two audiences: people, and the software that reads pages for them. On Cloudflare’s network, automated requests (search crawlers, monitors and scrapers as well as AI agents) now outnumber human ones.
If AI does the buying, do businesses still need a website?
Yes, when your business needs a source it can maintain and control. AI assistants can use a website to check your services, hours, policies and evidence. They can also use directories and third-party pages. Owning a site does not control the order an assistant searches or guarantee that it cites you. It gives you a place to publish an accurate record.
The evidence on where AI gets its information is more nuanced than either side of this argument admits, so here it is honestly:
- When the question names you, your site matters most. Across 6.8 million AI citations for local and branded questions, Yext found brands’ own websites were the largest single source, at 44%. Yext sells listings software, so read it as a vendor study.
- When the question doesn’t name you, other people’s pages matter most. On buyer questions like “who’s the best” that don’t mention a brand, Proofsource found only 0.8% of citations went to the brand’s own site, in a small October 2026 sample of nine brands.
So the website does two different jobs now. Other people’s coverage, reviews and mentions get you into the conversation. Your own website is where the AI checks the facts once you’re in it. Skip the first and you’re not shortlisted. Skip the second and you’re shortlisted with the wrong phone number.
Why can’t you rely on AI to describe your business?
Because you don’t control the answer, and the answers are often wrong or out of date. AI models summarise whatever pages they find, including old directory listings and third-party sites. Independent tests in 2025 found significant errors in close to half of AI answers about the news, and wrong contact details for about a third of brands.
| Study | What was tested | What went wrong |
|---|---|---|
| Seer Interactive, December 2025 | 178 “what is [brand]’s phone number?” prompts across seven AI models | 36% of the numbers given didn’t match the brand’s own customer service page |
| EBU and BBC, October 2025 | About 3,000 news answers from ChatGPT, Copilot, Gemini and Perplexity, in 18 countries | 45% had at least one significant issue; 20% had major accuracy problems, including outdated information |
| Tow Center, Columbia Journalism Review, March 2025 | 1,600 queries across eight AI search engines | Wrong more than 60% of the time when asked to identify a news source |
Notice what was happening in the Seer test. The AI answered almost every time, confidently, and a third of the time it was wrong. It wasn’t refusing. It was guessing from directories and old listings. That’s the risk for a business: not that AI ignores you, but that it describes you from somebody else’s out-of-date page.
You’ll never fully control what an LLM says about you. You can make the truth easy to find, consistent everywhere, and published on a domain you own, so the AI has something better than a stale directory listing to work from. Websites aren’t dead. They’re under-maintained.
What do the AI companies do with their own websites?
The companies building AI assistants still maintain public product pages, release notes and help centres. Those pages state what is available, where it works and what its limits are. The useful lesson is not how many updates they publish. It is that they keep a first-party record that customers can verify.
The public records include Claude Code’s release list, Anthropic’s newsroom and OpenAI’s API changelog. Many releases are small fixes, not big features. The value is a dated first-party record that a reader can check for current capabilities, availability and limits.
I use these tools every day, and when I need to know how one works, I don’t ask an AI and trust the answer. I go to the company’s own documentation, because a summary of last month’s version of a tool that ships nearly every day is a summary of a different tool. If the people selling AI answers send you to their website for the truth, that tells you something about websites.
What happens when something goes wrong?
A crisis makes the value of a first-party record clear. During the CrowdStrike incident in July 2024, the Cloudflare outage in November 2025 and the AWS outage in October 2025, each company used its own site to publish incident information. The links below let a reader check the company’s explanation rather than rely on a summary.
- CrowdStrike posted its statement and workaround on its own blog, which became a remediation hub. Australia’s cyber agency linked straight to it.
- Cloudflare published a detailed post-mortem from its CEO on the same day as its November 2025 outage, including the cause and that it was not an attack.
- AWS published its summary of the October 2025 outage on its own site.
I lived a small version of this in October 2026. On 21 October, PHP shipped an emergency fix for the flaws people called PHP2Shell. The PHP project and WordPress.org explained it on their own sites, and my team spent the next couple of days applying PHP updates across our clients’ sites. I wrote up what PHP2Shell means for WordPress sites on this site within two days, because a lot of the early advice got the most important point wrong: updating your plugins did nothing for this one. The sources everyone trusted were the ones on the PHP project’s and WordPress.org’s own domains.
When something new happens to your business, such as a recall, a price change, a new service or a rumour, an AI model can take weeks to catch up. Your website can say it in an hour. That’s your media. Nobody else owns it.
Does every small business need a website?
No. Some small businesses win work through referrals, Google Business Profile or social media and do well without a website. The ABS reported that 55.1% of employing businesses in its survey scope had their own website or app in 2024-25. That combined measure cannot tell us a website-only rate or describe all nonemploying sole traders.
| Employing businesses in ABS survey scope | Own website or app |
|---|---|
| All businesses in survey scope | 55.1% |
| 0 to 4 staff | 44.6% |
| 5 to 19 staff | 68.7% |
| 20 to 199 staff | 90.1% |
| Construction businesses in survey scope | 39.8% (40.9% use social media) |
Look at the jump. In this survey, 44.6% of businesses with 0 to 4 people employed reported a website or app, compared with 90.1% of those with 20 to 199. That matches what I see. A plumber with more work than hours may not need a website at all. A growing firm that wants to be chosen by people it has never met almost always does.
The website is becoming a premium asset rather than a commodity. The cheap brochure site that nobody updates is losing its point, because an AI summary or a Google Business Profile does that job. The website that earns its keep in 2027 is the one that publishes the truth, carries the proof, answers the questions buyers and agents ask, and turns a visit into an enquiry. I’ve written more about why small businesses need a website and when they don’t.
Who will visit your website in 2027?
More software than people, if the forecasts hold. On Cloudflare’s network, automated requests already outnumber human ones, although that count includes search crawlers, uptime monitors and scrapers, not just AI agents. Cloudflare’s CEO expects AI agents to drive much more of it by 2027. An agent researching for one person may visit far more sites than that person ever would, looking for the same thing: the truth.
- Automated requests passed human ones. In June 2026 Cloudflare data showed 57.4% of requests to a selection of sites on its network came from bots. That is a share of requests, not of people or buyers, and the bot count covers all automation, not only AI agents. Cloudflare’s CEO called the data “a bit messy”, and it covers Cloudflare’s customers, roughly a fifth of the web.
- One agent, many sites. Matthew Prince, Cloudflare’s CEO, said in March 2026 that where a person might visit five websites, their agent “will often go to 1,000 times the number of sites that an actual human would visit”. Read that as his illustration of where agents are heading, not a measured average.
- Some AI referrals convert well. Adobe’s July 2026 analysis found AI-referred visits to US retail sites had a 60% higher conversion rate than non-AI visits. That is an observed difference in US retail, not a forecast for Australian B2B enquiries.
I expect more of the comparison work to happen before a person reaches your site. Fewer people will browse ten sites to compare suppliers. Their agents will, and then the person will visit one or two, ready to decide. Both visits matter. The agent decides whether you’re on the list. The person decides whether you get the job. I built PostCliq to help marketing teams find revenue leaks after the click.
How do you build a website that agents and people can both use?
Start with readable, useful content and a working path to enquiry. Google’s current AI search guide requires pages to be indexed and eligible for a search snippet. It also requires the site to be included in Search generative AI features in Search Console. Include is the default, but an inherited exclusion can change that. No special AI file or schema guarantees selection.
Google says special AI files are not required for its search features. Ahrefs found 97% of valid llms.txt files in its sample received no requests in May 2026. That is evidence about requests in a selected analytics panel, not proof that every agent ignores the format. WebMCP remains an experimental interaction mechanism. I would prioritise a working service description, contact path and permission flow before adding optional protocols.
- Publish the facts an AI gets wrong about you: phone number, address, hours, services, areas, price ranges where you can, and policies. Same wording everywhere.
- Give news a home. A simple updates or news page, dated, so there’s one place to say what changed.
- Make every important task work with a keyboard. Use real headings, labelled controls and clear focus states. Keep key facts available as text and make expandable sections operable.
- Add structured data for your organisation, location and services, matching what the page says.
- Check that AI agents can reach you. Some firewalls and CDNs, Cloudflare included, now block AI crawlers by default. Decide on purpose what you allow.
- Ask the AIs about yourself every month and fix the source of anything they get wrong.
That’s what I mean by an agent-ready website: a website that an AI agent can find, read accurately, trust and act on, without a person having to interpret it. For the bigger picture of why agents are arriving so fast, I’ve written about personal AI agents becoming the next big industry.
Frequently asked questions
Will AI replace websites?
No. AI assistants and agents read websites to answer questions and complete tasks, so they depend on them. My expectation is a different audience: more agent research and some human visitors arriving closer to a decision. That is a prediction to test, not a result every site will see. Websites that are clear and accurate become more valuable, not less.
Do I need a website if I have a Google Business Profile?
Not always. If you’re a local trade or service with steady word-of-mouth work, a well-kept Google Business Profile and reviews may be enough. If you want to be chosen by people who don’t know you, publish detailed services or prices, or control your story, a website does what a profile can’t.
How do I stop ChatGPT getting my business details wrong?
You can’t control it directly, but you can fix its sources. Publish correct details on your own website in plain text, keep directories and profiles consistent with it, and remove or update old listings. Then check the AI’s answer again after a few weeks, because models and their search indexes update at different speeds.
What makes a website readable by AI agents?
Start with clear headings, readable facts, labelled controls and a contact path that works. Browser agents can use screenshots, the DOM or accessibility information, depending on the system. Google describes these different approaches. Accessible HTML helps people and gives agents clearer controls, but an actual task test is needed to show that a particular agent can complete the journey.
Is a website still worth the money in 2027?
For a growing business, yes, if it’s built and run as an asset: kept up to date, readable by AI and fast for people. A brochure site that never changes is worth less every year, because AI summaries do that job. The value is in the truth you publish and the enquiries it turns visits into.
