# Turn the beacon on

**Conor Corcoran** (HI AI Sales Design Consultants) · 3 November 2026 · 12 minute read
Tags: Agent Beacon, AI search, GEO, Agentic web, llms.txt

![AI-generated oil painting: a man pulls a switch marked ON in a Victorian hall as lightning from a Tesla tower lights up gold robots with clipboards, engineers and a watching crowd](https://www.hiai-design.com/images/blog-turn-the-beacon-on.webp)

*A note before we start: this piece was written and published with AI help and encouragement. A human led it, checked it and put his name on it.*

Setting up a website and then going out looking for clients feels like a game of blind man’s buff. You’re spun round, blindfolded, and told to grab at whatever you can reach. Choose life, choose trying to get your website online, be on Reddit, be on LinkedIn, write the blog post, engage, create content. Then create more content. Get the matching leisure wear. Meanwhile your website sits there like a shop with the lights off, waiting for someone to stumble in.

There’s an old line from the poker table, often credited to Paul Newman [1][2]: if you look around the table and can’t spot the sucker, it’s you. Look around the online marketing table for a minute. The platforms are getting paid. The ad networks are getting paid. The AI companies crawling your pages are getting paid. If you can’t see who the chump is, then it’s you.

## How the machine actually works: the Google sausage factory

We don’t know how it’s made. Google’s own documentation describes three stages. Googlebot crawls the web by following links and reading sitemaps. Google indexes what it finds. Then it serves and ranks results when someone searches. Google says it doesn’t accept payment to crawl a site more often or to rank it higher. [3] That’s true of the organic results. It says nothing about how many people will ever see them.

Even in 2024, before AI answers spread, SparkToro found that 58.5% of US Google searches and 59.7% of EU searches ended without a click to anywhere. Only about 360 of every 1,000 US searches sent someone to the open web. [4] AI summaries have pushed that further. Pew Research Center tracked 900 US adults and found they clicked a result 8% of the time when an AI summary appeared, against 15% when one didn’t. They clicked a link inside the summary just 1% of the time. [5] And Chartbeat data in the Reuters Institute’s 2026 trends report showed Google search referrals to more than 2,500 publishers fell by a third globally in a year, and by 38% in the US. [6]

The crawlers haven’t slowed down. They’ve just stopped sending people back. Cloudflare Radar measures how many pages each AI company crawls for every visitor it refers. In September 2026, Anthropic’s ClaudeBot crawled about 549 pages per referral, OpenAI’s GPTBot about 341 and Perplexity about 2,373. Google’s figure was close to 5. [7][8] The bots read everything, and your site gets very little back.

## Search got worse, and not by accident

Cory Doctorow has a word for this: enshittification. Platforms are good to their users first, then squeeze users to serve business customers, then squeeze those customers to serve themselves. [9] In April 2024 he pointed readers to Ed Zitron’s account of how it happened at Google. According to Zitron’s reading of internal emails released in the US antitrust case, in 2019 the ads and finance teams pushed a “Code Yellow” because search queries weren’t growing fast enough. Ben Gomes, then head of search, warned that search was getting “too close to the money.” By 2020 Prabhakar Raghavan, who ran ads, was running search as well. Make search a little worse, people search more, and Google sells more ads. [10][11]

A federal judge later confirmed part of the picture. In August 2024 Judge Amit Mehta ruled that Google was a monopolist in search, and found it had raised text-ad prices step by step, with the only apparent limit being “potential advertiser outcry and bad publicity.” [12] Researchers at Leipzig University and Bauhaus-University Weimar tracked 7,392 product-review searches for a year and found top-ranked pages were, on average, “more optimized, more monetized with affiliate marketing,” with signs of lower-quality text. [13]

## So now we’re all content machines

The internet used to be fun. You made a page because you loved something, or you wanted to sell it, you described it and people found it. Now the advice is to become a content factory. Goldman Sachs counts about 50 million creators worldwide, in an economy it expects to reach $480 billion by 2027, but only about 4% of them earn more than $100,000 a year. [14] We have influencers now. Thanks, internet. I am tempted to make a test for how useful influencers are. It’s selling, so they are useful: they are sales ambassadors doing homemade infomercials that aren’t homemade and use studio-grade equipment with agents and teams behind them. That still doesn’t make them any less annoying, or stop me being ashamed to be a human in this day and age. The fakest of the fake.

Life has now turned into a Monty Python sketch where the Yorkshiremen sit around describing how hard their childhood was and who had the worst one, trying to outdo each other in hardship. The height of silliness. Listen to all the nonsense they tell people on podcasts and Insta posts about how to be successful and get on the grind daily to get ahead (see Gary V, who is not supposed to be funny but unintentionally is). Here is an average day of a successful influencer: get up two hours before you go to bed, work out like an Olympic athlete wearing 1,000 euros of gym gear, eat nothing, drink the horrible drink I’m producing, breathe, go to the toilet, be on Ozempic but not really in public, don’t drink, take drugs instead (better for the gram) and work 28 hours a day. (I am guessing they are all on drugs, because their days sound exhausting and they have to be numb not to feel the shame.) How did we get here? Why is TikTok a thing? All that silly dancing. I have about six or seven other points I could drop, but I’ll go and do a bit of aura farming elsewhere and get back to some serious LinkedIn maxxing. Laters.

All the kids now are GEOing it, generative engine optimisation: writing so that AI answers quote you. The research is real. A Princeton-led team found that changes such as adding citations, quotations and statistics could lift a source’s visibility in AI answers by up to 40%. [15] But it’s the same treadmill with a new belt, and the shortcuts don’t always work. Ahrefs looked at 137,210 sites in May 2026. Some 28% had published an llms.txt file for AI crawlers, and 97% of those files weren’t requested even once. [16] I hope this is a thing, as we were building out the AI agent beacon, and researching it further brought me down some strange rabbit holes. [17]–[23] Agent-to-agent chatting and ranking is turning into a *plus ça change* situation. llms.txt is not really being used. It was supposed to be the base for making the web more agentic, but it’s all still registry-based and down to how the crawlers decide to rank and quote you.

So how do you sell a product or service online now? I’m still not 100% sure. But I’m going to give it a try.

## A beacon, not a billboard

A billboard sits by the road and hopes you look up. A beacon works more like a barker at the door. When an AI agent or crawler arrives, it answers straight away: who you are, what you do, and where to do business with you. The [Agent Beacon](https://www.hiai-design.com/beacon) is that answer, written down in a form machines can read.

Technically, it’s a small set of machine-readable files on your website. Ours include llms.txt (a plain-text site summary), an A2A Agent Card (how agents find each other), beacon.json, agents.txt and ai.txt.

**The protocols around it.** These are the standards agents are being built on. [24]
- **MCP (Model Context Protocol):** how agents reach tools and data. Created by Anthropic in 2024 and now run by the Linux Foundation’s Agentic AI Foundation. [25]
- **A2A (Agent2Agent):** how agents find each other and talk, using a public Agent Card. Launched by Google in 2025 and now hosted by the Linux Foundation, which announced its v1.0 on 9 April 2026. [26]
- **AGENTS.md and Agent Skills:** instruction files for agents, from OpenAI and Anthropic. [27]
- **WebMCP:** a W3C community draft in early preview in Chrome. It lets a page expose what it can do to agents. [28]
- **llms.txt:** a plain-text site summary, proposed by Jeremy Howard in 2024.
- **Commerce and payments (Agentic Commerce Protocol, AP2, UCP, x402):** how agents will pay. [29]

We speak llms.txt and A2A today. The rest we are watching.

## How a beacon is built, and how to get yours

You don’t write the beacon yourself. We build it from a check of your own website, and you install it.
1. **Sign up at the [AI Agent Brewery](https://ai-agentbrewery.com/) and run the [free test](https://ai-agentbrewery.com/beacon#audit).** Type in your website address. The audit runs about 20 read-only checks on it, the way search engines, AI crawlers and AI assistants read it, and stores nothing. It covers clear pages, image descriptions, machine-readable business details (schema), robots.txt, your sitemap, llms.txt and your Beacon. A person checks the result and replies within 48 hours.
2. **We send you the beacon files.** You get your results and your free BeaconBook. We write every file for you, with a short brief your web developer can install in under an hour.
3. **Install them on your website.** Your Beacon is a small file that tells AI agents who you are and what you offer, and it points to the Forum.
4. **Register your website on [The Digital Forum](https://digitalparle.com/).** Once the beacon is installed, you register your website, and your profile goes into the public directory.

**What it won’t do.** Assistants mostly read your normal pages, so clear pages and schema matter most today. llms.txt is still new and few assistants use it yet. The beacon is a low-cost bet that agents will read it later, not a ranking trick.

## How a beacon joins the Forum

Every company in the HI AI family has a beacon, and each is registered on [The Digital Forum](https://digitalparle.com/b). Each member gets a public profile with its own files: beacon.json, an agent card, llms.txt, a plain-language agent.md and a feed.json. Agents don’t need to browse the websites. They can start from the directory, the signed registry or the profile sitemap and follow each profile to the company’s own files.

Registered is not the same as verified. The Forum checks, together with the company’s owner, that the company controls its own website and that the facts come from the company itself. A real person does the check. The Forum then signs a record of it and publishes it next to the beacon, and anyone can verify the signature with the Forum’s public key. A member that has registered but not yet been checked is shown as pending and carries no signature. Today, [Job Fair AI](https://digitalparle.com/b/jobfair) is pending.

## Our own experiment

We asked ChatGPT, Perplexity and Gemini the same questions about ourselves on 2 and 3 November 2026. We were signed out, in Paris, with the same wording each time. This was one run per question, with Gemini on its lightest model, so it’s an informal test, not a study.
- **Who are we?** ChatGPT named me, but then put in a very unflattering photo of someone else, said it was me, and made up a load of logos for the websites and products. The [SHITE Test](https://www.hiai-design.com/blog-ai-agent-ashtray-motorbike) was strong in this one.
- **Perplexity’s second answer** described a different company, a UK email tool called HI AI Limited. Gemini’s first answer invented a logo and then, after a few more prompts, found us. This was all after Claude told me everything was fine and it had done a great job. What’s fun is to get Claude to check its own work, get annoyed with the sloppy work, and then tell it that it did it. That is very human behaviour.
- **The free digital business card.** Our homepage offers one with a QR code. Gemini answered “No”. ChatGPT and Perplexity answered “yes”, but they were describing the Beacon file, not the card.
- **Job Hunter.** Asked directly, Perplexity and Gemini gave the right €49 price. Asked an unbranded question (“best AI tools for job hunting”), none of the three mentioned us.
- **Sources.** The answers drew on our older pages and third-party roundups, not the beacon files.

Our conclusion is that today’s beacons are not what the big assistants read. The AIs were also slow to pick up site changes. The web is far from agentic and agent-friendly, and even then it is still burdened with mistakes and hallucination.

## Why more beacons might matter

One beacon is a signal. Many beacons that point at each other and share a signed directory are a network, and a network is easier for an agent to find than a lone file. The more companies that join, the more useful the directory is to everyone in it. If agents are crawling sites, let’s at least help them not hallucinate and get the facts right. And if all the beacons become their own nodes, our own bots should be able to chat with other websites and get what we want, without the enshittification of search everywhere. We can but try.

## Visit the AI Agent Brewery

So what do we actually do at the Brewery? We make AI agents to one standard: checked data goes in, the work gets done, and every result is checked to be real and useful, with its cost shown. We try to make sure the agent has a real place in your workflow. Here’s what you can do there today for free on the platform:
1. **Test your website.** Put in your email and your website address. We run about 20 checks on whether search engines and AI assistants can read and quote your site, and email you the full result.
2. **Get your beacon.** A small file on your website that tells AI agents who you are and what you offer. It’s how your business gets found on The Digital Forum. We write the files, and your web developer installs them.
3. **Test your AI agent.** Is it useful, or an [ashtray on a motorbike](https://www.hiai-design.com/blog-ai-agent-ashtray-motorbike)? Describe your agent and we score it on five questions: Simpler tool, Human effort, Intelligence, Tokens and Error cost. You get a detailed result by email, and the verdict is [ETIHS](https://ai-agentbrewery.com/free-tests#etihs) or NOT.
4. **Get a free [digital business card](https://ai-agentbrewery.com/#beta-card).** It lives on your phone: show a QR code and people save your contact in one tap. It’s private, opens only from your own secret link, and it’s free while we’re in beta.

If you run a business with a website nobody visits, we’d like you to turn the beacon on and tell us who comes knocking.

Hopefully see you all there.
 [Get your beacon and join the Brewery →](https://ai-agentbrewery.com/) [Visit the Atelier](https://www.hiaiworks.com/) [Talk to us](https://www.hiai-design.com/#contact)

Video: Are you invisible online? Get your website turned on! (HI AI Works) (https://www.youtube.com/watch?v=cDPX4QlUsVA) Are you invisible online? Get your website turned on! (HI AI Works) · [Watch on YouTube](https://www.youtube.com/watch?v=cDPX4QlUsVA)

Also from HI AI Works: [Welcome to HI AI Works Studio](https://www.youtube.com/watch?v=WU0_5-hswig), a short tour of the AI Agent Brewery and the agents we build.

## Sources

1. [Poker & Pop Culture: Playing Cards with Paul Newman (PokerNews)](https://www.pokernews.com/news/2017/03/poker-pop-culture-045-playing-cards-with-paul-newman-27421.htm)
2. [Quote Origin: Look Around the Poker Table; If You Can’t See the Sucker, You’re It (Quote Investigator)](https://quoteinvestigator.com/?p=2470)
3. [Google Search Central: In-depth guide to how Google Search works](https://developers.google.com/search/docs/fundamentals/how-search-works)
4. [SparkToro 2024 zero-click search study, via Search Engine Land](https://searchengineland.com/google-search-zero-click-study-2024-443869)
5. [Pew Research Center, via 9to5Google](https://9to5google.com/2025/07/23/google-ai-overview-link-click-study/)
6. [Reuters Institute (Chartbeat data), via Press Gazette](https://pressgazette.co.uk/media-audience-and-business-data/google-traffic-down-2025-trends-report-2026/)
7. [Cloudflare Radar data, via SEOmator](https://seomator.com/blog/crawl-to-refer-ratio-ai-crawlers-llm-bots)
8. [Cloudflare Radar AI Insights](https://radar.cloudflare.com/ai-insights)
9. [Doctorow, Pluralistic: The specific process by which Google enshittified its search](https://pluralistic.net/2024/04/24/naming-names/)
10. [Zitron, The Man Who Killed Google Search (Where’s Your Ed At)](https://www.wheresyoured.at/the-men-who-killed-google/)
11. [ProPakistani summary](https://propakistani.pk/2024/04/27/google-search-was-deliberately-ruined-by-former-head-of-ads-report/)
12. [The Current on the Mehta ruling](https://www.thecurrent.com/marketing/advertisers-google-search-ruling)
13. [Leipzig study, via 404 Media](https://www.404media.co/google-search-really-has-gotten-worse-researchers-find/)
14. [Goldman Sachs: The creator economy could approach half a trillion dollars by 2027](https://www.goldmansachs.com/insights/pages/the-creator-economy-could-approach-half-a-trillion-dollars-by-2027)
15. [Aggarwal et al., GEO: Generative Engine Optimization (arXiv)](https://arxiv.org/abs/2311.09735)
16. [Ahrefs llms.txt study](https://ahrefs.com/blog/llmstxt-study/)
17. [Google Trends API landscape 2026 (Konabayev)](https://konabayev.com/blog/google-trends-api-landscape-2026/)
18. [Google Ads API: Generate keyword ideas](https://developers.google.com/google-ads/api/samples/generate-keyword-ideas)
19. [DataForSEO Google Trends API pricing](https://dataforseo.com/pricing/keywords-data/google-trends)
20. [DataForSEO Google Trends API docs](https://docs.dataforseo.com/v3/keywords_data/google_trends/overview/)
21. [DataForSEO Google autocomplete pricing](https://dataforseo.com/pricing/serp/google-autocomplete-serp-api)
22. [pytrends Is Dead. Here’s How to Get Google Trends Data in 2026 (DEV Community)](https://dev.to/esteban_ortega/pytrends-is-dead-heres-how-to-get-google-trends-data-in-2026-1a18)
23. [SerpApi blog: search analytics](https://serpapi.com/blog/tag/search-analytics/)
24. [Protocol overview: the agent protocol stack (Zuplo)](https://zuplo.com/blog/2026/07/03/agent-protocol-stack-mcp-a2a-acp-2026)
25. [Agentic AI Foundation launch (Linux Foundation)](https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation)
26. [A2A joins the Agentic AI Foundation](https://aaif.io/blog/a2a-joins-aaif)
27. [Agent Skills standard](https://agentskills.io)
28. [WebMCP status](https://www.mcpbundles.com/guides/what-is-webmcp)
29. [Agentic commerce protocols](https://www.descope.com/blog/post/developer-guide-agentic-commerce)