This week in AI Search News: ChatGPT Ads open across Europe today, Reddit disappears from ChatGPT's citations for the second time in a year, AI crawlers turn out to ignore JavaScript links, AI Overviews cover half of French search, and Google starts building tools instead of linking to yours.
ChatGPT Ads open in 31 European markets today
OpenAI announced on 19 August that ChatGPT Ads reach 31 European markets from 24 August. Germany, France, Spain, Italy, Sweden, Norway, Denmark, Austria and the Netherlands are all in. OpenAI started testing ads in the United States in February and had reached eight further markets before this. Tens of thousands of marketers have now run ads on ChatGPT.
Ads only appear for people on the Free and Go plans. Plus, Pro and Enterprise stay clean. Buying is not self-service yet: campaigns run through OpenAI's Ads Solutions team, agency partners and technology partners, with an Ads Manager promised later. Targeting is geo plus custom audiences. OpenAI says advertisers never receive user conversations, and that advertising does not change the answers ChatGPT produces.
Digiday reported the consent side of the launch on 18 August. OpenAI's new EU privacy policy goes live later this month, and the European product is built on consent rather than on the profile-based targeting most media buyers expect.
Our view: two lanes are opening at once and most brands are not ready for either. The paid lane arrives in Europe today with almost no targeting to hide behind, so the ad has to match what the person is asking in that moment. The organic lane is the one that decides whether ChatGPT already names you when the ad is not there. Buying attention inside an answer that recommends a competitor is an expensive way to lose. Check what ChatGPT says about your category from a clean logged-out session before you brief anything.
Reddit fell out of ChatGPT's citations for the second time in a year
Promptwatch measured Reddit's share of ChatGPT Search citations falling from 3.83% to 0.52%, reported on 19 August. That is an 86.4% drop in four days, between 14 and 17 August. There were two steps down, a smaller one on 8 August and a sharp one on 14 August.
Nobody in the trade press could explain it. Promptwatch said plainly that it can identify when the shift happened, not why, and Search Engine Journal went back over it on 19 August without landing on a cause. ChatGPT's use of the site: operator jumped to 16.8% of fanout queries on 8 August, but the six-day gap before the real drop makes that a poor fit. Promptwatch could not rule out a problem in its own data collection either.
What we measure ourselves
We ran the same question across four client projects, 55 tracked prompts and two markets, on ChatGPT Search plus five other engines. That is 2,917 responses and 4,384 sub-queries, and the full analysis is here. Comparing 18 July to 7 August against 14 to 24 August, Reddit's share of ChatGPT Search citations fell from 4.01% to 0.13%. That is 96.9%, steeper than the number that got reported.
No other engine moved with it. Google AI Overviews went up 17% over the same window, AI Mode up 5%, and Perplexity more than doubled. This is not a fault in Promptwatch's data collection, and it is not the open web deciding Reddit matters less. It is one platform.
The mechanism shows up in the sub-queries. ChatGPT largely stopped doing broad open-web retrieval and started running domain-scoped site: searches against 86 specific destinations. Reddit is not one of them. Wikipedia is the check on that reading: it went from 0.88% to zero in the same window, which is what you would expect if general-purpose sources were being swapped for structured, category-relevant ones.
It has happened before. In September 2025 Reddit's share of ChatGPT citations fell from roughly 7% to roughly 1%, then recovered to around 3% within weeks. This is the second disappearance in eleven months, and the first one reversed.
Our view: reprice Reddit, do not cancel it. The first collapse came back, this one may too, and Reddit is still being cited by Google AI Overviews, AI Mode and Perplexity while ChatGPT ignores it. What deserves the freed-up attention is your own domain authority on the questions buyers actually ask: comparison pages, real pricing detail, specification tables. Those are the structured, category-relevant destinations ChatGPT moved toward. And measure share of voice, never citation counts, because a raw count falls whenever a model simply cites less.
AI crawlers found none of the JavaScript-linked pages
Vinicius Stanula ran a 41-day test on a Brazilian business directory and published it on 19 August. He split 1,062 pages into two groups. One group was reachable through plain HTML links. The other was reachable only through links injected by JavaScript.
Eight non-Google crawlers were watched: GPTBot, ClaudeBot, Bingbot, PerplexityBot, Meta-ExternalAgent, Amazonbot, ChatGPT-User and OAI-SearchBot. Across 41 days they discovered zero of the JavaScript-linked pages. Not a low number, none. Googlebot reached 2% of the JavaScript pages against 5% of the HTML ones. GoogleOther, which renders but does not build the search index, found 48% of them.
Google's rendering crawler sees your JavaScript navigation, so a rendering test looks fine, while the crawler that feeds the index and every AI crawler outside Google see nothing. Stanula also reports that recovery after launch was harder than getting it right first.
Our view: this is the cheapest audit in AI search and almost nobody runs it. Turn JavaScript off, load your site, and try to walk to your money pages using only the links that appear. Whatever you cannot reach that way does not exist for GPTBot or ClaudeBot. In the audits we run, broken discovery is far more common than weak content, and it is much faster to fix. Server-side render the navigation and treat client-side rendering as the exception.
AI Overviews cover more than half of French search
SE Ranking studied 100,000 French keywords across 20 niches and published the results on 18 August. AI Overviews appear on 52.63% of those searches, one month after launch.
That is roughly double the pace of the earlier European rollouts. When AI Overviews reached Germany, Spain and the Netherlands in March 2025, first-month coverage was 27.86%, 33.23% and 26.56%. France started where those markets took a year to arrive.
YouTube is the largest source by a wide margin, cited in 51.31% of answers and holding 14.33% of all citations. Facebook follows at 3.51%, then service-public.gouv.fr at 1.62%, Wikipedia at 1.33% and Instagram at 1.13%. Half of the ten most-cited domains are social or video platforms. About 90% of cited sources are French-language pages, and 60.46% of cited pages already rank in Google's organic top ten. Business and Relationships queries trigger an AI Overview more than 86% of the time. Fashion and Beauty is lowest at 20%.
Our own sweep of European business communities on 2 August found France was the only market with live, findable demand for AI visibility help, and every fresh French complaint thread postdated the AI Overviews launch there on 22 July. A 52% coverage rate one month in explains why. Dutch and German business owners were not asking, because in March 2025 only a quarter of their searches changed.
Our view: if you sell in France, the local-language requirement and the organic overlap are the two numbers to act on. Nine in ten cited sources are in French, so a translated page is not optional, and six in ten cited pages already rank in the top ten, so classic ranking work still carries most of the load. The YouTube share is the part that surprises people. A category where half the answers cite video, and you have no video, is a category you are absent from.
Google is building the tool instead of linking to yours
Google's generative UI reached AI Overviews this week, after launching in AI Mode. Search Engine Journal covered it on 21 August. When the model decides a query calls for it, Search now generates an interactive thing on the spot: a calculator, a simulation, a visualisation, a practice quiz. The user never leaves.
Google's own head-to-head numbers are striking. Generated interfaces beat the top organic result in 90% of evaluations on the main prompt set, and 73.5% against 19% on information-seeking queries. Against purpose-built websites they did less well, winning 35.3% while the custom sites won 50%. It is live globally in English in AI Mode, reaching around a billion users there and more than 2.5 billion in AI Overviews.
Our view: this is aimed straight at the utility page. If your traffic depends on a mortgage calculator, a unit converter, a dosage table or a formula tool, Google can now produce a working version of it inside the result. Finance, health, education and SaaS are the most exposed. What survives generation is what generation cannot reach: your own data, an authenticated workflow, a saved history, something transactional at the end. A page whose only asset is arithmetic is a page Google no longer needs.
Also noted
• The llms.txt spec shipped a v2 with formal Markdown link relations on 17 August, so agents can find the Markdown version of a page instead of guessing at the URL.
• Microsoft Clarity added an AI Scrape-to-Referral Ratio, comparing what AI crawlers take from your site against what they send back. It puts a number on a complaint publishers have only been able to describe.
• Googlebot now applies a single pass of HTML unescaping when reading JSON-LD. Double-escaped structured data that Google used to quietly repair will now fail, so revalidate any schema your CMS generates.
• One year of AI Mode data shows queries running about three times longer than classic search queries. People ask full questions there, which is why a page that answers in its first paragraph beats one that builds to a conclusion.
That is the week. If you want to know what any of this means for your own AI visibility, our door is open.