Welcome to this week’s Pulse: Google’s explanation for lost clicks tested; a case study showed why Core Web Vitals patches target the wrong thing; a report measured how agents managed pricing pages; John Mueller weighs in on agent access; and a longtime Bing leader announced his retirement.
This is what matters to you and your work.
AI insights don’t just absorb low-value clicks
Researchers shared new data from a randomized field experiment on Google’s AI insights. They found no measurable difference in bounce rates, search returns, or time on site between clicks made with summaries and those made without them.
Highlights: The study measured a 39.8% drop in organic clicks when AI previews appear. Losses were concentrated on informational queries, while navigational and transactional queries showed no measurable change on smaller samples. Liz Reid, Google’s vice president of research, said AI insights reduced “bounce clicks,” low-value visits that users quickly abandon, but did not release data to support this.
Why it matters
If AI insights primarily absorbed low-value visits, the extra clicks websites get when summaries are removed should look worse, but that’s not the case. Added clicks had the same bounce rates, dwell times, and return-to-search behavior as the others.
This leaves Google’s defense of click quality unsupported by experiment data. A drop in clicks on AI Overview queries cannot be considered as a simple loss of visitors who would not have converted. We covered the original experiment back in Aprilwhen first measuring click loss. Reid repeated the explanation of bounced clicks in several public contexts, and Sundar Pichai addressed the same traffic issue in Maybut Google still hasn’t released the segmented click data that would help fix the problem.
Read our full coverage: Google AI Overviews study reveals lost clicks were not of lower quality
LCP patches often target the wrong item
Google’s John Mueller highlighted a case study on why so many Largest Contentful Paint (LCP) fixes fail to improve the score. In store layouts that vary by merchant, the browser may lock on the wrong element, so that each subsequent optimization targets something that was never the LCP.
Highlights: The case study, published on web.devtraces a year of work on Core Web Vitals. The team traced the store’s low LCP scores back to the browser locking on the wrong element on the page, a side effect of the way its models loaded, and then adjusted the pages so that the browser measured the actual main content. The retailer reports that a higher share of its online storefronts subsequently adopted LCP.
Why it matters
The useful part here is the order of operations, more than the specific fixes. Before compressing another hero image, check which element the browser actually counts as LCP, because in templated layouts or with a lot of carousel it may not be the one you think.
Ours look at Core Web Vitals on CMS platforms found the same pattern in HTTP Archive data, where the real-world LCP tends to break when the platform is slow to allow the browser to discover the main image, not just when a page is heavy.
What people say
The post sparked a busy thread on Mueller’s LinkedInwith over 60 comments, and most of them focused on whether faster pages actually generate conversions.
The doubts fell into two camps. Regarding the causal link, Manhal Abou Zaki, SEO manager at Omnicom Media Group, described the link as indirect:
No matter how many case studies are presented, a faster loading page can certainly support and facilitate conversions, but it is unlikely to convince someone to convert just because the page loaded faster.
Francisco Antonio Fuentes Figueroa, an SEO consultant, weighed in on the reported conversion increase himself, noting that it was a full-year pre/post comparison with other changes delivered alongside and “no visible control groups,” and asking if there was any A/B or staggered rollout data to isolate the LCP patch.
David Swinstead, who publishes as The CRO Standup Comic, said he advocated for the quick conversion link, but was still hesitant about its size: “But a 9% increase? Counterbalance came from Georgi Petrov, founder of Uxify, who said his own 50/50 tests were the other way around, namely that “LCP changes move conversion more than people expect.”
The second camp took place in the classification. SEO Manager Vijay V. questioned whether Core Web Vitals were a ranking factor, saying he had “seen no evidence that they are.” Mueller’s response summed up the debate in one line: “Sure, I can make more money, but will I rank higher?” » Vishal Parmar supported the doubt with field observation: sites with poor LCP, CLS and INP that rank well for competitive conditions.
Others embraced the broader point made by the case study. Waqar Abdullah said the real lesson is to treat the LCP element as business critical, not just to make the site faster. Amarachi Kalu described it as “revenue rather than ticking the box”: conversions are “what businesses remember”. Mat Bennett kept it light, noting that the only place he still encounters layout changes is “in Google products.”
Read our full coverage: Google’s Mueller explains why LCP fixes miss the mark
AI agents turn to third-party sources when pages fail to load
A report from Siteline compared a simulated Claude agent to 100 of the top B2B software products to see if it could find their prices. When the agent encountered access errors or rates they couldn’t read, they often pulled the numbers from third-party sites.
Highlights: Siteline founder David Kaufman asked the agent to try to report prices and features on a set of leading B2B software sites. In a notable number of attempts, the agent encountered an access error or unreadable prices, and in these cases they were much more likely to abandon the brand’s own page and pull figures from external sources that may be outdated. Most failures were due to prices loaded in JavaScript that agents did not render, or prices hidden behind a sales contact.
Why it matters
An AI agent is now a visitor that you can accidentally lock. A page that appears complete in a browser may be considered empty by an agent if the important parts are loaded client-side after the initial fetch. When this happens, the agent does not stop. He finds the answer elsewhere, often on a third-party site or with a competitor who displays his prices clearly.
The solution is to make prices and key features readable in the initial HTML rather than loading them afterwards, and to highlight them on the page.
Read our full coverage: AI agents struggle to read B2B prices, report says
Mueller says don’t blindly block agent browsers
John Mueller was asked if Google’s quality principles would change as AI agents crawl sites for users. His response was that most of the principles are sound, with one new best practice worth noting.
Highlights: The question asked of Mueller on Bluesky was whether guidance such as “images provide a satisfying experience” still applies when the visitor is an information agent rather than a person. Mueller said he expects most of the principles to remain the same, because a site that is useful to users is usually also useful to agent browsers. He added that some details will evolve and cited not blindly blocking browser agents as a new best practice that will come into play.
Why it matters
As my colleague Roger Montti points out, Mueller’s response draws a line between two different things. Content quality standards are not being rewritten for the agentic era, because those standards were written for humans and agents serve those same humans. What’s new is technical accessibility. A site can meet all the quality requirements and still harm itself by blocking agents who increasingly act on a user’s behalf.
Roger compares this to the early days of nofollow, when some sites sculpted their internal links to accumulate PageRank and ended up starving important pages. Blocking agent browsers could follow the same pattern, a technical decision made for a single reason, leading to unintended visibility costs later. The takeaway here is to check your bot and your access rules before they quietly cut off a class of visitors. Mueller says don’t block blindly.
Read our full coverage: Google answers question about SEO for AI agents
Fabrice Canel of Bing announces his retirement
Fabrice Canel, a senior product manager who led Bing’s crawling and indexing team and championed IndexNow, has announced his retirement from Microsoft after nearly 30 years. He shared the news in a farewell message on LinkedIn, writing that he had accepted Microsoft’s voluntary retirement program, effective July 1.
Its release removes one of Bing’s most recognizable contacts for the SEO and webmaster community, at a time when Bing’s index quietly powers AI search products like Copilot and ChatGPT web results. This role is why Bing index is still important, even if its share of consumer research remains little.
Read our full coverage: Longtime Bing search leader Fabrice Canel retires from Microsoft
Theme of the week: Plumbing in research
The AI Overviews experiment aims to measure whether traffic generated by Google has ever been of high value. The LCP case study is also a measurement story, to see if the browser is rating the item you think it is. The Siteline report is about access, meaning whether an agent can read your page well enough to send a buyer to your site. Mueller’s response concerns the rules for this access and the cases where blocking agents turns into a visibility problem. Canel’s retirement is the human version of the same story, one of the people who built and explained this layer by moving away.
Overall, the week is less about a new feature and more about the underlying machinery, how sites are measured, read, and reached as more of the audience arrives through an AI layer. The work of remaining visible moves from one level to the next, whether the systems presented to your readers can use what you publish.
Top stories of the week:
More resources:
Featured Image: sparkling/Shutterstock
