Press review

Press review 14.11.2026: Google traffic attribution, an incrementality multiplier, a bot policy

A brand named in the first ChatGPT query appears in the final answer 68.9% of the time. A brand pulled in only during retrieval makes it 2.1% of the time. The rest stay out of the answer, however good their SEO is.

Bartosz Bielawski

Bartosz Bielawski

· 13 min read
Share
Press review 14.11.2026: Google traffic attribution, an incrementality multiplier, a bot policy

Google traffic landing in Direct? Check your channel rules, because the referrer carries only the host

On its Search Central site, Google published a full list of the referrer values its search traffic sends. The referrer contains only the scheme and host, for example https://www.google.com/ - no path, no parameters and no query.

That explains why some Google visits land in Direct or Referral in your analytics. The cause is your own channel definitions, referral exclusions and filters, not a change on Google's side. Discover and Google Images send the same values, so the referrer will not separate the surfaces for you.

Some visits will arrive with no referrer at all: private mode, a new tab, a mobile widget. On top of that, Search Console counts a click, while analytics counts a visit only after the page and the script have loaded, and ad blockers and consent banners block that script.

What to do: compare your channel rules with Google's list and check whether you are counting requests for images, CSS and JS as traffic. Separate Discover from Search using the Discover report in Search Console, not referrer data.

Sources:

Share of voice no longer tells you whether AI recommends your brand

Classic share of voice divides your brand's organic traffic for a set of queries by all available traffic for those queries. That formula assumes clicks and stable rankings, an environment that no longer exists: 68% of Google searches end without a click.

Large brands report tiny shares of traffic from AI. Booking Holdings and Etsy are below 1%, and LegalZoom is around 3%. At the same time, 94% of business buyers use AI in their buying process, so the assistant's answer increasingly replaces a visit to your site.

So the question is not how much demo traffic comes from organic, but how often AI recommends you when a buyer is ready to buy. Short prompts like "best boxing gloves" do not reflect how buyers describe their situation.

What to do: build a list of real questions from sales conversations and run them many times a month. Measure visibility, narrative control and commercial impact, not just clicks.

Sources:

A brand named in the first query almost always makes it into the AI answer

Models add well-known brands when they break a question down into sub-queries. A brand present in those queries appeared in the final answer 68.9% of the time, while a brand pulled in only during retrieval made it 2.1% of the time.

A Seer and Trustpilot study shows how much weight trust carries: brands without a review profile had a median citation rate of 1%, while brands with an active profile had 75%.

Most AI citations come from sources the brand neither owns nor pays for. Brands recommended by AI also received noticeably more visits within a week, mostly through searches that included the brand name.

What to do: treat brand presence in other people's content and in review profiles as your ticket onto the shortlist. Tactics built for AI only work after that stage, not instead of it.

Sources:

Robots.txt is voluntary - your CDN sets the bot blocks anyway

Robots.txt is only a declaration. Real enforcement happens in the CDN, the WAF and bot management services. Those settings can contradict robots.txt without the SEO team knowing, because they are configured on the infrastructure side.

Your site is now visited separately by search crawlers, AI training crawlers, retrieval bots, user-initiated fetches and autonomous agents. Blocking a training crawler has different consequences from blocking a bot that pulls content into an answer.

The author expects agents to look more and more like people, which will make them harder to tell apart. Blocking an agent that compares products can cut a brand out of the very answer someone uses to choose a supplier.

What to do: replace your block list with a bot policy. Write down which categories of bots you let in, then check whether CDN and WAF rules are quietly blocking them.

Sources:

AI can cite your article and still not recommend your company

A title about an SEO audit does not mean the text explains how to run an audit. Often it spends five paragraphs explaining why audits matter and never answers the question. What it is, how to do it and who to hire are three different questions, and each piece should handle just one of them.

The sentence "in this article you will learn how" is a promise, not an answer. The author advises building texts out of fragments that make sense on their own: the self-sufficiency of a paragraph is underrated. Instead of "we are the best", state a concrete limitation of your offer.

The key distinction: AI can cite your article without recommending your company. Citation and recommendation are two different things, and an inconsistent brand description across your site, directories and old publications damages both.

What to do: review the texts that are supposed to bring in enquiries and check whether each paragraph answers a question on its own. Then compare the company description on your site, in directories and in your Google Business Profile.

Sources:

Build a content cluster from the offer page, not from the blog

A cluster that works both in Google and in AI answers has three layers: a page for buying-intent queries, a pillar guide and supporting articles. The usual mistake is dozens of texts written without a plan, with no links to offer pages and with topics that compete with each other.

Order matters: first the offer page with a form, then a pillar guide that links to it early in the copy, then the supporting articles. Their number ranges from 5 to 30, depending on how broad the topic is.

You measure results at the cluster level, not per article. What counts is the offer page ranking on page one, organic and AI conversions on that page, and non-branded clicks in Search Console.

What to do: list your service's use cases and turn each one into a separate cluster. One complete cluster does more than 50 disconnected articles.

Sources:

Target frequency can finally mix non-skippable with Shorts and In-feed

Google updated its help page and now allows you to mix skippable, non-skippable up to 20 seconds, bumper, In-feed and Shorts formats in a single Target frequency campaign. Previously, ticking non-skippable automatically switched off In-feed and Shorts, as the October archive of the page confirms.

Mixed-format campaigns can only be bought at auction, because reserved buying applies only to unmixed formats. The recommended non-skippable length has increased to 7-20 seconds, and the frequency range is 2-7 per week or 4-12 per month.

The UK and Australian language versions of the page still say you cannot mix formats, and DV360 allows non-skippable up to 15 seconds only in an in-stream-only format. You cannot convert an existing campaign: you have to end it and create a new one.

What to do: in a new campaign, tick non-skippable, then check that In-feed and Shorts are still ticked. Prepare separate video assets: vertical 6-60 second videos for Shorts and horizontal 15- and 6-second videos for the feed.

Sources:

Trim a Smart Bidding budget with value rules, not bid adjustments

Pausing a campaign or leaving it limited by budget is not a neutral state: it costs you missed opportunities, not just money. Cutting delivery is a rookie move; it is better to optimise for a higher return on a smaller spend.

Smart Bidding respects value rules, unlike ordinary percentage bid adjustments. With a value rule you can go as low as 0.5, half your usual bid. Excluding a segment entirely is allowed; a partial adjustment is not.

Instead of pausing a whole Shopping or Performance Max campaign, pause a couple of the weakest ad groups for a few days. Raising a ROAS target works with a lag because the impact is smoothed out, while value rules take effect fastest.

What to do: when you overspend mid-month, use a value rule with a lower multiplier and exclude your weakest regions. Set a reminder to revert the changes so you start the new month on full settings.

Sources:

The Outlook outage may have lowered Microsoft Ads spend - do not blame your campaign

Microsoft Advertising warns that the Microsoft 365 outage may have affected ad spend in Outlook, even though the advertising platform itself stayed healthy. The incident started on 12 November at around 12:15 GMT and mainly affected Exchange Online.

The company distinguishes between platform availability and the availability of advertising opportunities. People who could not open their inbox generated fewer ad requests, so ads could still be delivered while some inventory was simply unavailable.

The practical consequence: a drop in spend and impressions during that window does not necessarily mean a configuration or bidding problem. Whether the incident is closed depends on the wider Microsoft 365 recovery status.

What to do: if you run Outlook campaigns, mark 12 November in your reports and do not draw conclusions about creative fatigue from it. Check whether you changed bids during that period, or you will credit the outage with a decision of your own.

Sources:

An incrementality multiplier shows how many platform conversions are real

The most common reason incrementality programmes fail is that test findings never reach the weekly reports that drive decisions. Most advertisers stick with first- or last-touch attribution even after a test reveals a channel's true impact.

The multiplier is simple: incremental conversions from the test divided by platform-reported conversions in the same window. A multiplier of 0.25 means only one in four reported conversions was actually caused by the ad, and a value above 1.0 means the platform undercounts.

You need the multiplier per funnel stage, not just per channel. In the example, LinkedIn scored 1.3 on demo requests and higher on qualified opportunities. Remarketing is mostly non-incremental, while prospecting is mostly incremental.

What to do: give untested channels a range instead of a default of 1.0 - in the example, the Meta campaign got 0.5 with a range around it. Put the multiplier into the weekly report, or the test will stay in a slide deck.

Sources:

Microsoft Advertising rebuilt its reports - there is now a dedicated bid strategy report

Microsoft is rebuilding the interface after complaints about navigation, reporting and account access. The new menu groups features by common tasks, breadcrumbs have been added, and tables now have better filters, date-range comparisons and custom columns.

Filters now support OR logic, show active filters and segments, and can be saved as custom views. The search terms report gained source labels, extra fields and faster access to negatives, and AI Max reporting got its own improvements.

The most useful addition is Bid Strategy Reports: average target CPA and ROAS in one place, target versus actual, trend lines and conversion delay information. Microsoft also admitted that some account recommendations were too generic.

What to do: during your next Microsoft Ads account review, open the bid strategy report before you change a target. Some features are still rolling out, so not everyone will see them yet.

Sources:

AI and tools

Claude joins Google Meet as a participant and writes your meeting notes

Claude joins video calls as a visible participant, for now only in Google Meet. You can ask it questions out loud during the call, and every participant can edit the documents and whiteboards it creates.

After the meeting, Claude writes the notes and action items, and each meeting becomes its own page you can return to with questions. It saves the transcript, summary and screen description, but not the audio. The feature is available on Team and Enterprise plans and is on by default, though in Enterprise an admin must enable it.

Claude stays out of calls with people from outside your organisation and leaves if such a person joins, unless an administrator allows it. Notes, transcripts and in-call answers are free during the beta, and documents and whiteboards are free until 10 February 2027.

What to do: if you record what was agreed with a counterparty, check your organisation's policy before you let an assistant into the call. Answers that draw on connected tools count towards your usage.

Sources:

Claude now has sheets and whiteboards on every paid plan - without pivot tables

Sheets and whiteboards in Claude are available on all paid plans, not only the most expensive one. The sheets lack pivot tables, dynamic tables and macros, and sharing is limited - it is more of a worse version of Google Sheets.

In tests, Claude's sheet and a plain Google Sheet with a Claude extension delivered virtually the same results. The whiteboard is more interesting: it works well at the ideation and planning stage, especially if you think visually. Dashboards can produce visual explainers focused on specific data.

For comparison, similar spaces in ChatGPT are only available on the $100 plan, and Claude has no native image generation. Separately, DOTS users finally got custom instructions of up to 35,000 characters.

What to do: do not move spreadsheet work into Claude if you rely on pivot tables. Use the whiteboard to plan a campaign or a content structure, then move the output into the tool your team actually works in.

Sources:

An AI agent in production: iteration limits, a budget and human approval for risky actions

A simple agent breaks as soon as real people use it, because it cannot handle situations outside the script. Stuffing the whole company policy into the system prompt makes it hard to maintain, and an agent without tools to fetch context will invent an answer or promise a refund.

A better pattern is a knowledge base in files, skills that point to the right file, and workflows that let you test individual parts without running the whole agent. Force the answer into JSON and validate it, because one extra line of text breaks the next step.

Agents run in a loop: when a tool returns something unexpected, the agent keeps trying, and every round costs money. You set a maximum number of iterations, a cost cap per session and a monthly budget with an alert at 80%. A refund above 75 euros pauses the workflow until a human approves it in Slack.

What to do: before an agent touches customer support, cut off its direct access to database queries. The workflow should give it only the relevant person's data, and risky actions should sit behind an approval gate.

Sources:

Dreambeans shows how Google assembles personalisation from five of your sources

Dreambeans, a Google Labs experiment, is available to all US consumer accounts for users over 18, on Android and iOS, with no Google AI subscription needed. The app builds content from five sources: Search, Workspace, YouTube, Photos and Gemini chat history.

Generation runs through a pipeline: context, topics, stories, selection, images. Personal context is materialised per date, and every story carries a subject, a rationale, a category and a confidence level. The "Why am I seeing this" screen shows how many signals were combined into a single recommendation.

The part that matters for businesses: source data and derived content have separate lifecycles. A change in the source app shows up in new stories after 24-48 hours, but cards that have already been generated stay until someone removes them manually.

What to do: read this as a preview of personalisation based on context rather than clicks. If someone asks you about privacy in products like this, show them exactly that gap between the source and the derived content.

Sources:

A YouTube channel makes a video in 40 minutes and pulls in $10,000 a month from it

A channel with simple stickman animations published 25 videos in three weeks, roughly one a day, fully automated. Three weeks after launch it was getting 12.5 million views a week, and one video reached 10 million views in 12 days.

The whole process runs through Claude Code and one tool that generates the animation world. One video took 40 minutes and cost under 12 credits, about 1% of a five-hour allowance. Estimated ad revenue is around $10,000 a month.

The author stresses that what wins is a consistent visual world, which cheap imitations lack, combined with the structure: simple graphics and a short loop that keeps people watching to the end.

What to do: if you are considering a Shorts channel for a brand, work out the cost of one video first and plan a series, not a one-off. The format only works if you publish daily and keep the visual style consistent.

Sources:

Analytics, marketing and e-commerce

GA4 has an AI Assistant tab - check how much revenue assistants bring you

Google Analytics 4 has an AI Assistant tab that shows traffic, sessions and completed goals coming from AI assistants. Store examples show that even 10-50 thousand visits from AI can generate meaningful revenue.

On lead generation sites, the same thing shows up in form submissions: conversions from AI traffic arrive in large numbers but get lost in the overall channel report. Without separating that traffic, you cannot tell whether growth came from SEO, Ads or assistants.

That single tab is the simplest way to show a store owner what AI traffic is really worth, without building your own attribution model.

What to do: open the AI Assistant tab and record a baseline for revenue from assistants. Then track it monthly, because the channel is growing, and without this tab it will vanish into Direct.

Sources:

Reddit is testing product review summaries - you cannot buy your way in

On 10 November, Reddit started testing Reviews: summaries of product discussions, recurring pros and cons, and links to the source threads. The test covers the US and English, about 10,000 products in four categories.

Purchases are not verified, so a summary may be based on opinions from people with no experience of the product. Reddit itself warns that AI can miss context and that sentiment only reflects the conversations in that pool - it is not a percentage of satisfied customers.

Brands cannot pay to have their products included in Reviews, and there are no conversion results to point to yet. Reddit plans to add more categories and expand beyond the US in 2027.

What to do: treat Reviews pages as another listening post, not a source of truth about sales. Look for patterns - if the same complaint comes up across different threads, that is a signal to change the product or its description.

Sources:

Facebook leads: a higher-intent form and one qualifying question filter out the noise

The default form type in Facebook Ads is More volume, which brings in more leads, often of poor quality. The Higher intent type adds a review screen where people confirm their details: fewer leads and possibly a higher cost per lead, but more people who pick up the phone.

A live form cannot be edited, only duplicated. The strongest filter is a question only a real customer would answer, such as the type of job or the city - that screens out job seekers.

The open locations setting is on by default and shows your offer to people interested in travelling to or buying in that place, not only to residents. Phone verification by SMS only works in the mobile app and requires the phone field to be mandatory.

What to do: in lead campaigns, check the form type and add a qualifying question before you raise the budget. Results depend on the market - in the example, a pressure washing company started winning commercial jobs worth about $20,000.

Sources:

Read more