Press review 20.11.2026: budgets in Ads, a sitemap and the price of a job
One advertiser cut Google Ads spending by 75 percent, while deleting blog posts lifted clicks by 14 percent even as AI visibility fell by 62 percent.
Trending topics
Topics that clearly more independent sources have covered in the last two weeks than before:
SEO and AI search
Deleting blog posts lifted clicks even though AI visibility fell
After a batch of blog posts was removed, impressions in Google's AI features dropped by 62 percent while organic clicks rose by 14 percent. That is not a measurement error - it is the difference between exposure and outcome.
The numbers are concrete: AI impressions fell from 1,535 to 591 a day, while clicks rose from 235 to 268. The blog produced 87 percent of AI impressions from 36 percent of pages, but its CTR in classic Google was 0.45 percent against 2.1 percent on commercial pages.
What to do: before you delete posts, compare AI feature impressions with clicks for the same URLs in Search Console. Content that buys visibility with no clicks is not a win - it is a cost.
Sources:
- Search Engine Land, 2026-11-19: Deleting blog posts hurt AI visibility, but organic clicks told a different story
AI Overviews answer 82.5 percent of brand queries - check what they say about you
AI Overviews appear on 82.5 percent of US brand-name queries (SE Ranking). On "best X for Y" queries they show up 96 percent of the time (Semrush). A model describes your brand more often than you think.
Models build a picture of a company from the whole web. An outdated About page or an inconsistent price list produces a wrong answer without any hallucination - the model simply repeats old data with confidence. A confident but outdated answer does more damage than no answer.
The sources disagree on where citations come from: Muck Rack says 84 percent come from earned media, Profound says almost 60 percent come from company websites. In practice you need both in order.
What to do: run "what is this company", "best X for Y" and "X vs competitor" in ChatGPT and Gemini, write down what the model says about price, offer and audience, then fix it at the source.
Sources:
- Search Engine Land, 2026-11-19: How to fix what AI search gets wrong about your brand
Google Search now shows a panel with your recent AI questions
Google is rolling out a collapsible sidebar on the desktop homepage globally. It shows recent AI searches and lets you pick up an interrupted thread.
Google explains the change by saying people ask multi-step questions rather than "one and done". The panel also lets you start new threads in AI Mode. The interface had been tested for several weeks.
What to do: watch organic CTR separately for desktop and mobile. Longer AI Mode sessions can mean fewer clicks to sites at the same impression volume.
Sources:
- Search Engine Land, 2026-11-19: Google rolling out a collapsible sidebar for Google Search
Add one line to the board report: referrals from AI assistants
A board report starts with money, not rankings. AI assistant referrals get their own line next to organic, paid and email - because leadership will ask what it earned anyway.
Take first-party data from three places: GA4 (sessions and revenue), Search Console (impressions in generative AI features, number of URLs) and Bing Webmaster Tools (citations, cited pages, grounding queries). The two AI views in GA4 - the AI Assistant channel and a custom grouping of ChatGPT, Claude and Perplexity - overlap, so keep them separate and say so in the report.
What to do: close the report with three sentences - what you could measure, where the brand shows up, what you will check next. No invented AI revenue.
Sources:
- Search Engine Land, 2026-11-19: Introducing SEO+ reporting
WordPress and technical SEO
WordPress 7.1.5 fixes 18 flaws - one lets an attacker take the server without logging in
The new release patches 18 vulnerabilities and two security improvements. The most serious is an unauthenticated privilege escalation in Template Part block attributes that leads to remote code execution.
With that kind of flaw an attacker reads and changes files, writes to the database, creates admin accounts and installs backdoors. On multisite, a flaw in Customizer menu previews lets an administrator take over the whole network. Separately, stored XSS via shortcode attributes in comments needs no login at all.
What to do: update every site you manage to 7.1.5 now, starting with multisite installs and sites with open comments. Check whether your host blocks automatic security updates.
Sources:
- Search Engine Journal, 2026-11-20: WordPress 7.1.5 Security Release Fixes 18 Vulnerabilities
Your sitemap is not broken - Chrome 158 stopped styling it
Chrome 158, rolling out since 17 November, removed client-side XSLT support. /wp-sitemap.xml now shows up as a wall of XML tags.
Search engines do not use XSLT - Googlebot and Bingbot parse raw XML. Indexing and rankings will not suffer as long as Search Console reports "Success" for the sitemap. It is a purely visual change.
What to do: warn clients before they report an outage. The fix arrives with WordPress 7.2 on 8 December, or with Yoast SEO 28.9, which restores a styled view using CSS. Until then, check sitemaps in Firefox.
Sources:
- Search Engine Journal, 2026-11-20: Your WordPress Sitemap Isn't Broken – Chrome 158 Has Dropped XSLT
WordPress 7.2 redirects to a random post - by default and with no switch in the admin
Version 7.2, planned for 8 December, adds a /random/ URL that redirects to a random published post. The feature is on by default and has no toggle in the dashboard.
Turning it off takes one line in a mu-plugin - and note the filter name listed in the November developer roundup does not match the name used in the code. The random pick also ignores noindex settings from SEO plugins, so posts you would rather hide can surface.
What to do: decide per site whether /random/ should stay before the update. On cached sites, purge the cache after the first admin visit or the URL may return 404.
Sources:
- Search Engine Journal, 2026-11-20: WordPress 7.2 Arrives Dec. 8 – Last-Minute Changes Still Being Made
Google Ads and paid media
Advertisers are cutting Google Ads budgets - one cut spending by 75 percent
Long-standing Google advertisers are scaling back search spending because rising costs and less prominent ad placements undermine the point of the campaigns. One retailer cut its monthly budget from $2,000 to $500 within a year.
AI Overviews push ads further down the results, and Google does not guarantee placement inside those answers - so you cannot predict delivery. Money is moving to Tubi, Bing, Reddit and Nextdoor, and one advertiser's costs tripled or quadrupled in three years.
Google says AI features drive more searches and that user experience comes before monetisation. AI Mode still generates less ad revenue per query than classic Search.
What to do: before you cut a budget, check whether the drop comes from cannibalisation by AI Overviews or from account quality. Those are two different decisions and two different fixes.
Sources:
- Search Engine Land, 2026-11-19: Advertisers cut Google Ads spending amid rising costs, AI changes
AI, tools and automation
Google is teaching models quietly - no version number, no announcement
On 27 October Google Research described a method for moving knowledge from short-term memory into model weights on a repeating cycle. The cycle has a "wake" phase, where the model takes in data, and a "sleep" phase, where it consolidates and learns from synthetic samples.
Consolidation gives new facts their own parameters instead of overwriting old ones. For now this is research on open Qwen models, not Gemini - but if it reaches production, the knowledge cutoff stops being a constant and becomes a schedule you cannot see.
The takeaway for brands is uncomfortable: what matters is not whether a fact appeared somewhere, but whether it is consistent across sources. The model learns the version that fits best, so conflicting versions of your offer become a risk.
What to do: keep one canonical, dated page of facts about the company and remove contradictions between your own site and the directories and profiles that describe you.
Sources:
- Search Engine Journal, 2026-11-20: Google's Next Update Won't Touch The Index And Nobody Will Announce It
A $30 model did the same job as a $142 model
Across seven tasks - from building a 3D world to an animated explainer - GPT-6.2 Astra cost $30 in total, while Fable 5.2 cost exactly $142. Both models charge the same price per token, so the difference does not sit in the price list.
It sits in token count and time: Astra delivered the tasks roughly five times faster. Fable won where image quality mattered, Astra where finishing the job mattered.
What to do: count the cost of a completed job, not the price of a token. In content production and simple pages the difference in the bill shows up faster than the difference in quality.
Sources:
- Jack Roberts, 2026-11-20: OpenAI Just Dropped a Dumb Genius (Astra 6.2 vs Fable 5.2)
Google shipped its third top-tier model in ten days
Gemini 4 Neon is Google's third frontier model in ten days. The hierarchy now has three levels: Carbon for code and complex work, Neon for deeper analysis, Hydrogen for everyday help.
Neon has a very long context window and an X-High thinking level, and accepts text, image, audio and video. One market curiosity: because Carbon is currently half price in a promotion, the model lower in the hierarchy costs more than the one above it.
What to do: do not switch production processes to a new model on launch day. Test it on one task type and compare the cost per job - names and prices change faster than your workflow.
Sources:
- Bijan Bowen, 2026-11-19: Gemini 4 Neon Is INSANE – Is THIS Google's BEST Model Yet?
Only 0.5 percent of Reddit mentions reach ChatGPT - quality beats volume
Of all Decathlon mentions generated on Reddit, only 0.5 percent made it into ChatGPT's sources. The rest does not exist for the model, even though it looks like a win in media monitoring.
A separate thread from the same webinar: after enabling consent to train models on YouTube channel content, the brand's AI visibility rose sharply. Larger creators more often switch that consent off, so smaller ones break through to the models more easily.
What to do: do not count mentions, count the ones a model actually cites. Pair web monitoring with the list of sources LLMs cite and check what really gets through.
Sources:
- Linkhouse Polska, 2026-11-19: Monitorowanie widoczności w AI | Chatbeat, Surfer, Peec AI, PromptEye | Webinar Linkhouse
Business and market
Adobe paid $1.9 billion for Semrush to get into AI optimisation
Adobe approached Semrush itself with a $1.9 billion acquisition offer. Semrush was not for sale - a few months earlier it had closed a two-year strategic plan and acquisition roadmap.
Adobe considered three or four companies. Semrush won on scale, data and an enterprise customer base. Adobe treats optimisation for AI search as one of the fastest-growing areas of martech, and sees smaller companies as riskier targets because they depend on founders.
What to do: when planning your 2027 stack, assume SEO data and AI citations will be sold in enterprise bundles rather than as standalone SaaS. Watch the pricing announcements.
Sources:
- Search Engine Land, 2026-11-19: Why Adobe bought Semrush: Former president reveals new details