Press review

Press review 06.11.2026: the homepage is no longer the health metric, and AI ads are counted in the hundreds of thousands

Three in four homepages lost organic clicks year over year, and on more than half of sites the homepage no longer collects the most AI impressions. Meanwhile, Adthena counted 175,000 paid placements in Google's AI answers, and ChatGPT Ads already has more than 100,000 advertisers.

Bartosz Bielawski

Bartosz Bielawski

· 11 min read
Share
Press review 06.11.2026: the homepage is no longer the health metric, and AI ads are counted in the hundreds of thousands

Topics that clearly more independent sources have covered in the last two weeks than before:

Three in four homepages are losing clicks, and on most sites they no longer collect AI impressions

Signal: 2 independent sources this week, 2 in 30 days; tracked since 2026-11-05.

In an analysis of 110 homepages across six industries, 81 sites, or 74 percent, lost organic clicks year over year. The median decline was 21 percent, with the middle half of the sample landing between -37 and +3 percent. SaaS was hit hardest: 21 of 25 homepages were down, with a median of -35 percent.

A second finding in the same material may matter more: on 57 percent of the sites studied, the page with the most impressions in Google's AI features was not the homepage. Branded queries accounted for 79.6 percent of homepage clicks, so the front page is no longer a discovery engine. It is becoming a confirmation point for people who already know the brand.

What to do: report branded clicks across the whole domain and conversions by landing page, not homepage traffic. A single homepage chart no longer describes the health of a site.

Sources:

A brand can appear in 91 percent of AI answers and still be described incorrectly two times out of three

Signal: 3 independent sources this week, 10 in 30 days; tracked since 2025-11-17.

The study covered roughly 2,000 scored mentions, five dental software brands and four engines: AI Overviews, Copilot, Perplexity and ChatGPT. One brand showed up in 91 percent of the answers where it should have, but only about a third of those answers got the customer profile, differentiators and product facts right at the same time.

The spread between engines is wide. Copilot invented a price in 153 of 223 mentions where no price could be verified, and attached unsupported negative descriptions to half of its statements. ChatGPT had the cleanest sourcing and the best customer-profile score (1.72 out of 2, against roughly 0.9 for the others).

Brand presence itself changed between runs of the same prompt in about a third of cases. So a presence metric tells you whether you appear, not whether you are described well.

What to do: separate the two metrics. Presence and accuracy are different questions, and the second needs its own set of control questions plus regular checks of what the engines say about price, differentiators and audience.

Sources:

The Search Console click spike between 31 October and 2 November was bots, not people

Some sites saw clicks spike 2 to 6.5 times while impressions and positions stayed flat. The spikes hit pages ranking 12 to 54 and produced CTRs of 45 to 100 percent across dozens of one-off queries, mostly on desktop.

One site saw desktop clicks rise 14.1-fold against its six-month median (from 21 to 297 a day) while mobile stayed flat. On another, a long-tail Canadian page jumped from zero to 40 clicks on 50 impressions, with the clicks coming from 23 variants of the same question using different numbers. There are four likely causes: scraping bots clicking links, CTR manipulation networks, AI agents doing query fan-out, and a gap in Google's bot click filter.

What to do: annotate your reports for 31 October to 2 November. Without the annotation, the return to normal levels will look like a traffic loss.

Sources:

A grocery chain outranked Wikipedia for "apples" and sold no extra apples at all

A small four-store chain in Connecticut published a "Types of Apples Guide" in April 2026. When the article was written, the guide ranked first in some US results, above the AI Overview for some queries and below it for others.

The owner said plainly that the rankings delivered "zero extra business", and the author never saw the store's Search Console data. The query is informational, geographically distant from the store, and leads to no cart. The ranking is real, but the intent does not match at all.

What to do: do not report rankings for purely informational queries as a win. Before one goes into a report, check whether the query has any path to a purchase or a lead.

Sources:

175,000 paid placements in Google's AI answers, and no setting to control them

Signal: 3 independent sources this week, 7 in 30 days; tracked since 2026-08-05.

Adthena counted roughly 175,100 paid placements in AI Overviews and AI Mode panels in the US between June and September 2026. Apparel accounts for about 111,000 of them and consumer electronics for 55,000, with the rest split across travel, beauty and toys. Paid presence in AI answers is concentrated in two categories.

There is no separate way to buy ads in AI Overviews or AI Mode. All Search, Shopping and Performance Max campaigns are automatically eligible, so you cannot target AI alone or opt out without switching off Search entirely. What remains are negative keywords and brand exclusions.

The scale is growing fast: AI Overview impressions in the US rose from 304 million in August 2025 to 575 million a year later, while they grew by 160 percent in the UK and the EU and by 536 percent in Australia.

What to do: when someone asks how much budget goes to AI, the honest answer is that it cannot be broken out in reporting. Say so before the question comes up, not after.

Sources:

ChatGPT Ads: 204,000 paid placements, more than 100,000 advertisers and 89 countries

Across five categories, the study counted 204,489 paid placements between June and September 2026. The split: apparel 90,517, electronics 56,569, travel 48,134, toys and games 4,694, beauty 4,575. The number of advertisers in the measured categories rose from 1,082 in June to 3,741 in July and 6,333 in August.

Query structure differs from Google. In travel, 69 percent of ads went to research and discovery, in toys 67 percent, in apparel 48 percent and in electronics only 42 percent. In beauty, the share reached 83 percent.

What to do: if you are testing this channel, start with "best" and comparison queries rather than branded ones. People who already know the brand are not the main audience here.

Sources:

From 9 November, ChatGPT Ads moves to Target CPA and Target CPC, and the old strategies disappear on 1 January

Target CPA replaces Max Bid in new conversion campaigns, and Target CPC replaces Max CPC in click campaigns. The old strategies stop working entirely on 1 January 2027. Autobid is being renamed intelligent bidding.

The system will adjust bids and delivery on its own, including bidding above or below the target. Existing campaigns keep running on the old model but can be switched manually. The vendor's warning is blunt: overly tight targets or a weak conversion signal will throttle delivery and leave budget unspent.

What to do: before migrating, check your signal quality (EQS, Pixel, Conversions API) and base your targets on your real CPA rather than on the budget you hoped to spend.

Sources:

Signal: 3 independent sources this week, 6 in 30 days; tracked since 2025-10-01.

All it takes is a conversion action with the same name in that account. Historically, a GCLID or phone number was accepted only in the account that generated the click. Today an upload with another account's GCLID goes through, and after extended correspondence Google Support confirmed that this is expected behaviour.

The result is inflated conversions in the Conversions and cost-per-conversion columns. According to Google, such conversions do not feed Smart Bidding, but the authors did not verify this independently.

There is a separate reporting trap: a "Successful" status in the interface and "#OK" in the upload history only mean the row could be read. In the example given, 100 uploaded conversions, 90 OK statuses and 60 conversions visible in the account amount to a real acceptance rate of 60 percent.

What to do: compare the number of uploaded conversions with the number visible in the account. A gap of more than 20 percent means the reported result does not reflect what actually happened.

Sources:

Be Active case: ROAS up 110 percent while spend fell nearly 60 percent

The initial problem was not budget but a rigid structure for a product range with no single season: gym, outdoor and garden, ski and bike accessories, pet products. The account was split into separate thematic pillars, with dedicated asset groups for seasonal categories and for promotions.

Low-margin products were excluded from promotions, shifting budget to high-margin items, especially the company's own brand. On top of that came a separate Performance Max campaign built only on text, image and video assets, without the main feed.

In peak season, conversion value rose 71 percent and ROAS more than 50 percent, while costs grew 10 percent. Comparing November and December 2025 with the previous year, ROAS was up more than 110 percent with spend down nearly 60 percent.

What to do: before adding budget, check whether the account structure matches the seasonality of the product range. Here the answer was not more money but a different split.

Sources:

AI and agents at work

The same rules scored 100 percent recall in AGENTS.md and just over 50 percent in a skill

The result comes from a Vercel evaluation that tested different ways of packaging rules for agents. The difference lies not in how well the rules are written but in where the agent looks: reading AGENTS.md files is built into the system prompt, while skill contents get skipped.

The pattern that worked is mundane: a short SKILL.md plus a folder of separate rule files (about 70 in Vercel's repo, each with a bad and a good code example and a note on performance impact), so the agent reads only what it needs.

A counterexample from the same discussion: adding rules and documentation to a code review skill made the results worse than the original. More instructions do not mean better results.

What to do: if your rules live only in a long skill description, check across a few tasks how many of them the agent actually applies. It is a cheap test, and the gap is often twofold.

Sources:

DeepSeek's super agent: a router, workers and a watchdog cut the slowest task from 29.5 to 15.7 minutes

The starting point was practical: a person can keep track of roughly three to five sessions at once, and a single session often never used sub-agents even when told to, doing all the work itself.

In the router-plus-workers setup, the router writes no code; it only hands out tickets. Each ticket has its own folder with requirements and verification criteria, and separate merger, checker, reproducer and tester roles each run in their own context. A watchdog runs a small script every 10 minutes and compacts memory once it exceeds 150,000 tokens.

One architectural curiosity: DeepSeek ran the router on a lighter model, judging its job to be simpler - the opposite of the usual practice of putting the strongest model at the top. Workers got Sonnet, because the ticket already spells out what to build.

What to do: if an agent gets lost on a big task, split the task into tickets with acceptance criteria before reaching for a stronger model.

Sources:

Nearly 40,000 stars: a workspace manager keeps Claude Code, Codex and Cursor sessions in one panel

The tool is open source and restores sessions after you close the terminal. Sessions can run on a remote server, because the server and client layers are separate. The interface is mouse-first, and on first install the tool detects which agents are present on the machine.

The most interesting part is a skill installed via npx skills add herdrdev/herdrskill. It lets an agent from one vendor query other workspaces and pass information between them, for example from a Claude Code session to Codex.

The problem it solves is real: in a bare terminal, closing the window kills every session, and parallel work quickly turns into chaos.

What to do: test it on one project before moving your daily work there. The value only shows up with three or more parallel sessions.

Sources:

A skill is a packaged prompt, and one in five skill ideas dies because there is nothing to automate

From the AI Tech Leaders team's experience: a skill pays off in two cases - either the task repeats daily or weekly, or it is rare but complex and needs a long instruction.

Evaluations are worth running only where the skill is critical, comparing changes on the same test cases before and after; session metadata produces false results in both directions. Treat other people's skills like executable files: scanners find data leak paths, but their scores are unstable. In production, the agent gets a read-only account, because it breaks data more often than code.

What to do: review your skills for the "used once a quarter" kind and remove them from auto-discovery. Every such skill takes up context on every run.

Sources:

Marketing and channels

Email delivers more visits than paid search, while organic has lost 13 percent in two years

Between August 2024 and August 2026, organic traffic fell 13 percent in the US and 24 percent in France. Over the same period, email visits rose 16.3 percent and 17.0 percent respectively, with email delivering 3.94 billion visits in the US against 3.70 billion from paid search.

Paid search grew 20.8 percent in nominal terms, but excluding the googleadservices.com domain the growth drops to 13.5 percent, so part of the increase is a measurement artefact. Direct traffic remains the largest channel, at 43.5 percent in the US, and is declining more slowly than organic.

There is also a measurement warning: a bug during the GPT-6 rollout in October meant links carried neither a referrer nor UTM parameters for about a week. Adobe measured roughly 31,000 fewer daily ChatGPT visits and 26,000 more Direct visits. Some of "AI took our traffic" is really "AI did not send a referrer".

What to do: whenever traffic declines, check whether the measurement changed before you blame the algorithm.

Sources:

AI visibility is closer to PR than to performance, and measurement has to reflect that

In Marie Haynes' sample, ChatGPT sessions amounted to 0.62 percent of organic traffic from Google, and research by The SEO Works found that organic search delivers 39 times more traffic than AI assistants. Kevin Indig could not establish a credible link between AI visibility and subsequent brand search or direct traffic.

Mike King summed up the debate with "the numbers are correct, but the category is wrong". The author's conclusion is not that GEO does not work, but that leadership wanting AI optimisation to behave like performance marketing does not make it a performance channel.

What to do: measure presence on real buying questions and ask customers about their first point of contact, then trace that through leads to revenue. Promising a return within weeks is overreach.

Sources:

Microsoft Clarity will show which AI agents visit your site and what they do there

The new agentic analytics module is available globally to all users, including those on the free plan. It shows AI agent traffic from searches and forms through logins and downloads to bookings and checkout. Simulations let you test whether an agent can complete a key task before a real customer fails at it, starting with commerce and travel.

Brand Agents are designed to help products get discovered in Copilot, Bing and MSN and to collect leads for local businesses. On WordPress, setup involves installing a plugin, choosing "Clarity with Brand Agents", publishing and connecting a CRM. Ask Clarity accepts natural-language questions about user behaviour and links its findings to session recordings.

What to do: create a separate segment for agent traffic in your analytics before it starts distorting conversion reports. An agent that completed checkout is not the same as a customer.

Sources:

Read more