Earlier this month we published the results of an experiment that failed on purpose, or at least failed instructively. AI agents built and ran 12 experimental news sites, and over two months OpenAI and Anthropic crawlers made around 115,000 requests to them. In the Gemini answers we sampled, not one page from the network was ever cited.
That raises the obvious follow-up question, and it is the one clients have asked us most since: if a steady stream of fresh content does not earn AI search citations, what does? This post is our answer, built on what the experiment showed us and on 18 years of watching search engines decide who to trust.
By Frank Vitetta, Founder, Orchid Box. Based on our own server logs and citation sampling, August to October 2026.
| Measure | Result |
|---|---|
| Sites built and run by AI agents | 12 |
| Logged period | 8 August to 8 October 2026, varying by site |
| Requests from OpenAI and Anthropic crawlers | About 115,000 |
| Citations seen in sampled Gemini answers | 0 |
| Google search clicks, one site, June to October | 20 |
| Share of AI crawler visits to English pages on multilingual sites | 93% |
Full methodology and caveats are in the original research write-up.
Short answer: No. Crawling only means a bot requested your page; a citation means an assistant chose your page as a source, and our sites never made that jump.
The most important lesson from the experiment is that AI search visibility has stages, and most sites confuse the first stage with the last.
Our network made it through the first stage and stalled, as the full server log findings show. Thousands of requests from training and search bots did not translate into a single observed citation, and Google Search Console recorded just 20 search clicks for one of the sites over four months. Volume of crawling told us nothing about trust.
Short answer: The same signals search engines have long rewarded: history, links and mentions from trusted sites, real readers, clear entities and original information.
No AI company publishes a citation algorithm, so anyone claiming to know the exact formula is guessing. What we can say is that the signals assistants appear to lean on look very like the signals search engines have always rewarded, which is exactly why a brand-new network could not fake them.
Established sites have years of crawl history, stable URLs and a consistent record of what they publish. Our sites had weeks. An assistant choosing between a ten-year-old trade publication and a domain registered last month has an easy decision to make.
Link building has not stopped mattering in the age of AI search; if anything, it matters in a broader sense. Assistants weight what others say about you: trade press, industry lists, review platforms and expert round-ups. Earning those mentions is what digital PR is for. A citation in a respected publication is a signal to Google’s index and to every model trained on or retrieving from it. Our network had no inbound links worth the name, and it showed.
Sites that people search for by name, return to and share build a footprint that automated content cannot simulate. Organic search performance and AI search visibility are now closely linked, because the assistants that ground answers in a search index inherit that index’s judgement about which sites deserve to rank.
Assistants need to know who is speaking. A company with a consistent name, a clear description of what it does, named experts and structured data and entity optimisation that tie it together is far easier to cite accurately than an anonymous site. Entity clarity is the foundation of every AI search optimisation programme we run.
Our sites repackaged information that already existed elsewhere, so an assistant always had a better, more authoritative source for the same fact. Content that contains something new, such as first-party data, a measured result or a genuine expert view, gives an assistant a reason to cite you specifically. There is a useful irony here: the post describing our failed network has attracted more attention than anything the network itself published, because it contains original findings nobody else has.
Short answer: Content that answers a specific question directly, contains evidence nobody else has, carries a named expert author and is promoted where trust already exists.
The experiment was a clean test of one content marketing strategy: publish a large volume of competent, sourced, AI-written articles and wait. It did not work, and Google has made its position on the approach explicit. It ran three spam updates in 2026 alone, in March, June and August, with scaled content abuse a main target throughout.
Content marketing that earns AI search citations looks different. It is slower and more expensive per piece, and it compounds:
AI tools are excellent for research, structuring and speed. The experiment showed they cannot substitute for the reputation that makes content worth citing.
Short answer: Combine server logs for crawler activity, sampled prompt tracking for citations and analytics for AI referral traffic; no single source shows the whole picture.
One practical finding surprised several clients. Google Analytics and similar tools run as a script in the visitor’s browser, and most bots never run it. Your analytics dashboard can therefore show nothing while AI crawlers request thousands of pages. In our 12-site experiment the only reliable record was the server logs.
Measuring AI search citations is harder still. Assistants answer differently each time, by country, by user and by phrasing, and most citations never produce a click. Any monitoring tool, including our own LLM Scout, can only sample answers. Sampling is still essential, because it shows trends in answer share and which sources are cited instead of you, but it should be read as a trend line rather than a census.
A sound measurement setup combines three views: server logs for crawler activity, sampled prompt tracking for citations, and analytics for the referral traffic that does arrive from AI engines. Our AIO analytics page explains how we combine them.
Short answer: Look for partners who refuse to guarantee citations, can read your server logs, invest in reputation over volume and explain the limits of their measurement.
The experiment also sharpened our view on what clients should expect from any agency selling AI search visibility, ourselves included. Shortcuts are being sold aggressively right now, and some of them are close cousins of the network we built to test. A few questions separate credible partners from the rest:
An AI search citation is a link or named source an assistant such as ChatGPT, Gemini, Perplexity or Copilot shows alongside its answer. It signals that the assistant drew on your page and trusted it enough to credit it.
There is no fixed timeline. In our experiment, brand-new sites were crawled within days but no citation appeared in sampled Gemini answers over the test period. Established sites with existing authority usually see movement within months of improving their content, entities and mentions.
Not on its own. Our network published continuously and was never cited. Volume without reputation, original information and trusted mentions did not earn AI search visibility, and scaled content is a target of Google’s spam updates.
No. Google Analytics runs as a script in the visitor’s browser and most bots never execute it. Server logs are the only reliable record of AI crawler requests.
It builds on SEO rather than replacing it. Assistants that answer from a search index inherit that index’s view of your site, so technical SEO, authority and content quality remain the foundation, with entity clarity and citation tracking added on top.
Our 12 experimental sites were crawled constantly and never trusted. That gap is the whole story of AI search right now. Assistants are not looking for more content; they are looking for sources they have reason to believe, and that reason is built the slow way, through history, reputation, original information and the endorsement of sites that are already trusted.
The good news for established businesses is that much of that trust already exists; it just needs to be made visible to AI systems. That is the work our AI search optimisation service is built around, alongside the B2B SEO foundations it depends on. If you want to know how often AI assistants recommend your brand today, and which sources they trust instead, get in touch for a free AI search visibility audit.
