Press review 21.11.2026: people vanish from the graph, poisoned rankings, AI labels in ads

Google's knowledge graph lost 10 percent of its person entities, and rank trackers are being served pages that look normal but list other sites. There is a simple test in Search Console.

Share
Press review 21.11.2026: people vanish from the graph, poisoned rankings, AI labels in ads

Google's knowledge graph dropped 10 percent of people - check whether your panel survived

In November 2026 Google rebuilt its knowledge graph. The number of person entities fell by 10.1 percent, while companies barely changed.

Most of what disappeared were duplicates: entries with the same first and last name, no subtitle and no description. Nearly one in five people with a score of 24 vanished overnight.

A Kalicube sensor that checks three thousand people and three thousand companies every day recorded a drop of about 6 percent on 16 November. Part of it returned by the middle of the next day.

What to do: stop reporting the entity score as a KPI. Check whether the panel shows up, whether the subtitle and description are correct, and clean up duplicates in your company data.

Sources:

Your ranking dropped but traffic is flat - that may be a poisoned result, not a fall

Since mid-September 2026 Google has been serving rank tracking tools result pages that look normal but list different sites than usual. The industry calls them poisoned SERPs.

Only automated queries are affected - manual searches look normal. On such a page the tracked site is about twenty times more likely to fall out of the top ten.

Reddit, Facebook, LinkedIn and Quora show up more often, and image packs appear nearly twice as often. AI overviews appear less often.

What to do: confirm every ranking jump in Search Console and with a manual search. Set an alert for mass disappearances of many URLs at once - that is the signature of poisoning, not a real drop.

Sources:

Google says AI expands the buying journey. Independent studies show fewer clicks

Google defends AI in search with an Ipsos survey it commissioned itself. Consumers using AI reportedly went through twelve touchpoints, almost three times more than the rest.

The Pew study says otherwise. With an AI summary, users clicked ordinary results in 8 percent of visits; without one, in 15 percent.

Seer Interactive calculated that organic clicks for informational queries fell by 61 percent over fifteen months.

What to do: report search clicks and AI citations separately. Also check whether the drop concerns informational or transactional queries - conclusions only differ there.

Sources:

Three in four AI citation losses are temporary - do not rewrite a page after one day

Ahrefs examined about 37,500 first-day citations across five AI products. Around half of the losses came back within three days.

Most cited sites appear for only one day; brands last longer. In Gemini 64 percent of answers carry no citation at all, in Copilot 53 percent.

In Google's AI overviews, AI mode and ChatGPT fewer than 1 percent of answers have no citation. There, losing a citation almost always means another site took the spot.

What to do: keep a daily citation log per question and report how many of the last seven days carried a citation. Review a page after two or three weeks, not after a single miss.

Sources:

Google tells marketing chiefs: fewer lists, more of your own data

On 19 November 2026 Google wrote from its Think with Google account that content, not AI, is what makes brands invisible. The tagline: stop doing “Top 10” lists.

The article linked in that post itself offers “Top 10 Things to Consider” as an example of good content. Neither the article nor Google's guide mentions self-promotional “best company” lists.

The Google author's actual recommendation is different: plan content around leads and sales, not around proxy metrics.

What to do: go through your list posts and check point by point whether any item repeats common knowledge. Replace those with your own data, tests and customer cases.

Sources:

A support number from an AI overview led a journalist to a scammer

Andrew Shotland called a Vanguard support number he saw in a Google AI overview. The person on the line posed as fraud detection, and the caller ID showed “Bank of America”.

The scam came out the next day. Neither Vanguard nor Bank of America had a record of those calls. Shotland is not reporting a search bug, only his own reaction: do not call support numbers taken from AI.

What to do: in finance and healthcare, check that the phone number, address and policy number are identical in the company panel, on the site and in external profiles. Mismatches make impersonation easy.

Sources:

Conversions rose 416 percent and the result got worse - check when your form fires

In one insurer account conversions rose 416 percent, impressions fell 32 percent and CPA dropped 21 percent. The numbers looked great; the leads were fictional.

The reason was simple: the conversion fired when someone started filling in the form, not when they submitted it. Automation learned to collect exactly those events.

In the same account, Google Search spend fell 49 percent while search partner spend rose 69 percent.

What to do: keep one, at most two primary conversions and bid only on them. After a landing page change, test live whether the event still fires where you think it does.

Sources:

A 20-dollar lead can cost more than an 80-dollar one - count the qualified lead cost

In Google Local Services Ads you pay for every qualifying enquiry. That pool includes spam, calls from outside your area and jobs you do not offer.

Existing customers calling about ongoing work also count as leads, even though they bring no new revenue. Spend 2,000 dollars with 20 percent unqualified enquiries and you waste about 400 dollars.

In Europe, automated credits for poor leads are not available. You can file feedback, but a credit is not guaranteed.

What to do: label the outcome of every enquiry - spam, outside area, unsupported service. Compare the cost of a qualified lead, not the cost of a lead.

Sources:

Ads now sit next to AI overviews in Bing - no new campaign, no report

Microsoft Advertising shows ads next to AI overviews in Bing. They run through your existing campaigns, with no extra setup, in every market except China.

Microsoft names Performance Max and AI Max as the main campaign types with access to this placement. It also says purchases after a Copilot interaction rise 21 percent - its own data.

No separate controls or reports for the placement have been announced, so measuring the effect will be hard.

What to do: check that Performance Max and AI Max campaigns run in your markets, and compare Bing separately from Google. A small-budget test will give you a firmer benchmark than Microsoft's announcement.

Sources:

Microsoft labels ads “Made with AI” - the declaration is signed by the advertiser

The label appears in the corner of supported ad images in Microsoft Advertising and in Bing results. It is meant to tell consumers the visual was created with AI.

AI content transparency controls are generally available since November 2026. Advertisers declare themselves whether an image or video contains AI-generated content and can pick a visible watermark.

Microsoft's detection is advisory, not decisive. The advertiser is responsible for a correct declaration.

What to do: inventory creatives built with generated images and declare them before the label appears on its own. Decide who signs that declaration - the agency or the client.

Sources:

Content, brands and trust

E-E-A-T appears 120 times in Google's guidelines - and increasingly decides citations

YouTube and Reddit are the two most cited domains in AI overviews. In an Ahrefs analysis from November 2026 they hold 24.5 percent and 18.1 percent of the top fifty domains.

In a Semrush study of 325,000 prompts, LinkedIn was the second most cited domain in ChatGPT Search, Google AI mode and Perplexity. 59 percent of the cited content came from people, not company pages.

Only 35 percent of Americans trust AI-generated content. In October 2026 Google added that fabricating author profiles is a form of deception.

What to do: build real author pages for experts, with portfolios and links to their profiles, then push them off-site - to conferences, LinkedIn, Reddit. The about page is often a client's most cited page and is usually underfunded.

Sources:

Salesforce's CEO said “Goodbye SEO”. Salesforce's own guide calls SEO the foundation

On 18 November 2026 Marc Benioff wrote “Goodbye SEO” on X while promoting Palmata, a tool for tracking brands in AI answers. Salesforce acquired Contentful, which owns the tool, on 1 September.

Salesforce's own SEO and AEO guide from 11 September says plainly that SEO is still the foundation. It recommends tracking both sets of metrics - organic traffic and AI citations.

SEO consultants answer the same way: optimising for AI rests on the same foundations and does not replace SEO. Others add that the market for AI monitoring tools is already saturated.

What to do: show rankings and citations side by side in your report, and count leads separately. “We are cited” and “we sell because of it” are two different metrics.

Sources:

A typical digital PR campaign is designed for journalists, and the links land on the homepage. Three months later the sales pages have not moved.

The author suggests reversing the order: first a map of commercial pages by revenue, margin and seasonality. Then a study that naturally points at your products, and only then a story for the media.

Her split: 60-70 percent of monthly links to priority sales pages, 20-30 percent to secondary targets. In one case a company earned 23 publications from flight cancellation data.

What to do: count the links pointing to each sales page, excluding the homepage. Pages with high margin and zero links are your brief list for the next quarter.

Sources:

AI does not search your keyword, it searches questions around it - check which ones you answer

One question in an AI search engine triggers several extra searches. Expanding a question into related topics is called query fan-out.

For a content author this means a page does not have to answer only the broadest query. It is enough that it explains one important detail of the decision well.

What to do: list the questions people ask while making a specific decision and check which ones you actually answer. Instead of publishing more similar articles, fill the gaps.

Sources: