A London car hire site lost 62 percent of its Google AI feature impressions after an agency deleted many of its blog posts in August, yet its organic clicks rose 14 percent. Search Engine Land published the figures in a contributor piece from Luca Tagliaferro, who runs LT Digital Agency, about his client G&M Direct Hire. It is one client, not a controlled test, so treat it as a field note.

The author is explicit about the limits. He says he cannot claim the deletions caused the click gain, because the site also saw other edits and seasonality shifts clicks. Comparing June 1 to Aug. 17 against Aug. 24 to Oct. 28, daily AI impressions fell from 1,535 to 591 while daily organic clicks went from 235 to 268.

The page-level export explains why the two metrics diverged. Across a year of data, 420 pages picked up AI impressions, and blog posts made up 36 percent of those pages but 87 percent of the impressions. A single Bolt vs. Uber pricing comparison produced 46,435 AI impressions, 25 percent of the site’s total. In classic search the same post drew 1.83 million impressions and 4,909 clicks, a 0.27 percent click-through rate.

A battery-failure warning post shows the same pattern, with 7,663 AI impressions against 65 organic clicks, despite close to 90,000 classic impressions. Across the site, blog pages collected 12 million organic impressions and 53,741 clicks (0.45 percent CTR), while commercial and service pages collected 3.3 million impressions and 69,142 clicks at 2.1 percent. Informational content was feeding AI surfaces. Pages that sell were producing the visits.

The author also tested Semrush AI Visibility tracking across May through October, benchmarked against four rivals such as Uber and Hertz. It reported a visibility score of 16 out of 100, with mentions down 10.3 percent to 26 and citations down 42.3 percent to 225. He found the topic clustering opaque: 241 prompts were grouped into 144 topics, most topics rested on a single mention, and the best-performing topic was tagged U.S. for a London business. Semrush owns Search Engine Land, and the contributor notes that nobody asked him to bring Semrush up. The vendor in question is the publisher’s parent, which is worth keeping in mind when weighing the critique.

Volatility is the deeper problem. The author cites SparkToro and Gumshoe.ai, who sent identical prompts close to 3,000 times. Their finding: ChatGPT or Google AI would give the same brand list twice with odds under 1 in 100. He also cites University of St. Gallen researchers, who measured just 34 to 42 percent agreement in cited sources from one day to the next when the prompt stays fixed. They put the standard error of a single run at 0.37 on a 0 to 1 scale and advise a minimum of seven daily runs of each prompt, sustained for two to four weeks. By the author’s own arithmetic, that works out to 78,400 monthly runs to cover 100 prompts on four AI tools.

For measurement, he suggests three layers. Exposure comes from page-level AI impressions in the Generative AI features report of Search Console, plus citation counts and share from Bing’s AI Performance report inside Webmaster Tools. Outcomes come from clicks on the same pages, branded search, and GA4 conversions. Demand comes from Performance report queries and Bing grounding queries, with a “How did you hear about us?” form field as a low-tech check. Third-party trackers, in his view, belong in research, not in KPIs.

The practical stake for a search team is attribution. A content cull that lowers an AI visibility score can look like a loss on a dashboard while revenue pages hold or gain. Before approving or reversing a pruning plan, split blog and commercial pages in Search Console, compare AI impressions against clicks and conversions for each group, and treat any single-run tracker figure as a sample, not a measurement.

Search Engine Land, contributor article by Luca Tagliaferro, published November 19, 2026.