The AI Visibility Brief: July 13-19, 2026

THIS WEEK

• Two AI agent studies this week: a website built for agents nearly doubled their task success / hidden prices are where they get stuck
• A review of 45 studies finds no technique that reliably gets content cited across AI platforms
• Websites that block AI training crawlers are still getting their ChatGPT clicks (Sorry we missed this one last week)

01

Two AI agent studies this week: a website built for agents nearly doubled their task success / hidden prices are where they get stuck

Two studies of AI browser agents, the tools that visit websites and act for a person, came out this week. The first, a controlled experiment Said Elnaffar and Farzad Rashidi posted to arXiv July 13, built two versions of the same online store with identical products and prices: one designed the usual way for people, the other built so an agent could read and act on it, with clean structure and plainly labeled facts and buttons. Across three different agent models, the agent-ready store took task success from 49% to 89%, with the biggest gains on exactly the jobs a shopping agent gets sent to do: pulling product details, comparing options, and choosing under several constraints at once. The second study came from Kevin Indig, who audited real business websites to find where agents give up, and the weak point was pricing. When a website hid its price, agents left to find the number on a third-party website in 45% of attempts, and even when the price was right there on the page they still went elsewhere 18% of the time.

We think this matters because agents are moving from searching for information to buying products and services on a customer's behalf. That is the shift to prepare for: when a customer sends an agent to compare and buy, the websites that win those purchases are the ones the agent can read and act on all the way through.

Businesses should act on this by making the pages where buying decisions happen as plain for a machine as they are for a person: facts stated in text, options and buttons clearly labeled, and deciding details like price and limits written where a machine can read them, rather than trapped in a calculator, a screenshot, or a PDF. That plainness is what took the experiment's agents from failing half their tasks to finishing nearly all of them.

02

A review of 45 studies finds no technique that reliably gets content cited across AI platforms

Olivier Martinez reviewed 45 studies published between late 2023 and this month on how content gets cited by LLMs, and posted the survey to arXiv July 15. It is one researcher's synthesis, not a new experiment, and its conclusion is blunt: no technique in the reviewed research reliably improves citations across platforms. The two levers that hold up are plain ones. Pages genuinely relevant to the question get used, and when a tool gathers material to write an answer, the sources it reads first get quoted more than the ones it reads last. A page cannot set that order directly; being the closest match to the question is what puts it near the front.

We think this matters because there is a lot of confident advice for sale about winning AI citations right now, and this survey is 45 studies' worth of reason to ask what evidence sits behind any of it. We build in this field, and the honest summary is simple: work that makes a page genuinely more useful for the questions buyers ask holds up. Few things do that better than real customer stories, since they carry the specific detail that comes straight from real work.

Businesses should act on this by putting real customer stories on the pages buyers actually visit: testimonials and write-ups that show the specific problem a customer brought and the work that solved it. The more specifically a page speaks to the question a buyer is asking, the closer a match it is when an AI tool goes looking for an answer, and that match is the one lever this research says holds up. A roofing company with a page walking through how it replaced a hail-damaged roof for a family on a deadline is exactly the kind of page these tools reach for.

03

Websites that block AI training crawlers are still getting their ChatGPT clicks (Sorry we missed this one last week)

Researchers at Bocconi University studied ten months of real US desktop browsing through a Comscore panel, following what people actually did after asking ChatGPT a question. Among the 3,844 websites the team classified, 79% blocked at least one major AI training crawler. Those blocking websites still received more ChatGPT referral visits than the websites that allowed everything, about 6.5 per website against about 4.5. The block and the traffic never connected because two different bots are involved. The training crawler reads the web ahead of time to build the model. A separate fetcher grabs a page at answer time, when someone's question needs it. A rule in robots.txt (the plain text file that tells bots what they may read) written against one does not bind the other.

We think this matters because it separates two decisions that usually get treated as one. Refusing to let AI companies train on your content is one choice. Staying reachable when a tool searches the live web for an answer is a different one, and the live side is where the citations and the clicks come from. For the many publishers who set that block, the AI traffic they may have written off is still arriving.

Businesses should act on this by checking which bots their website's rules actually name before assuming either outcome. The training crawlers and the live-search fetchers carry different names in robots.txt, and each can be allowed or refused on its own, so a publisher can keep its writing out of training models while still letting ChatGPT fetch a page when a reader's question calls for it. A block aimed at the wrong name does nothing, and a block aimed at the live fetcher is the one that actually removes a website from the sources AI answers can cite.

Thanks for reading this week’s newsletter.
Sincerely,
Brett