Press review 17.11.2026: a bot calls your business, revenue per session in ChatGPT Ads and a read-only agent

In a study of 102 sites, 84 of 87 advertisers earned less revenue per session from ChatGPT ads than from Google. Google counters that a single video in Performance Max lifts conversions by 12 percent.

Share
Press review 17.11.2026: a bot calls your business, revenue per session in ChatGPT Ads and a read-only agent

Google will call your business and book the customer - check this toggle in your profile

Ask Maps in Google Maps and Search now finds an open slot and calls the business on the customer's behalf. Appointment search is live only in the US and only in four categories: beauty, pet care, auto and home repair.

Ask Maps itself is already available in 30 European countries, so it will reach you sooner or later. Automated calls work in every US state except five.

Your business profile has a toggle called "Bookings and inquiries from customers" that turns these calls off. Google has not said whether Ask Maps respects it, so do not assume the switch is enough.

What to do: check the automated calls setting, fill in your hours and pricing in your profile, and warn whoever answers the phone that Google bots may call.

Sources:

ChatGPT citations drop quietly on the day a new model ships

After a new model version launches, ChatGPT citations can halve within a week. That is what happened on 9 July, when the GPT-5.6 family arrived: some publishers lost 50 to 90 percent of their citations.

This is not a penalty for your content. The new model picks sources differently - it leaned more on government and NGO pages and less on media outlets. Journalism's share of citations fell from about 33 percent in January to 22 percent in September.

On top of that, only 12 percent of links cited by AI rank in Google's top 10 for the same prompt. A good search position does not guarantee a place in the answer.

What to do: mark model launch dates on your citation chart. A drop with no content change and no decline in Search Console almost always means OpenAI changed something, not you.

Sources:

Search Console filters have reached the API - your reports can now run automatically

The Search Console API now supports filters that were previously available only in the interface: branded queries, multimodal search, and combined multi-country and multi-device filters.

The branded filter arrived in the UI about a year ago, the multimodal one in September and the country and device filters in October. Anyone pulling data through the API had no access to them.

In practice, you can now separate branded from non-branded traffic automatically, with no manual exports and no stitching spreadsheets together.

What to do: add brandedClassification and type=webMultimodal to your queries, then check the docs to see which filter combinations are supported before you rebuild your reports.

Sources:

AI reads a passage, not a page - so every section has to stand on its own

AI retrieval systems pull individual passages rather than whole documents. A paragraph taken out of context can end up in an answer with no clue as to which product or place it describes.

The takeaway for content is simple: every key section has to make sense on its own. Repeat the product name, location and conditions within the section, not just in the heading above it.

The second piece is a fact layer: approved product data, policies and evidence to back up your brand claims. Without it, the model has only your opinion, not a source.

What to do: take your three most important product pages and read a paragraph from the middle of each. If you cannot tell what it is about, rewrite it.

Sources:

A plugin update can be the attack - the version number proves nothing

Rank Math SEO PRO version 3.0.123 was distributed with malicious code. The severity was rated 9.8 on the CVSS scale, and the compromised file was served from the evening of 15 October until the next morning.

The problem: the same version number had shipped on 7 October as a clean file. One number, two different files - one safe, one not. A similar earlier incident affected WordPress.org plugins and about 35,000 installations.

For an SEO team, this is not purely an admin issue. Odd redirects, new URLs and changes to indexing directives are often the first sign of a compromised plugin.

What to do: keep an update log with the date, time and package source, and decide who reports suspicious indexing changes before you start looking for the cause in your content.

Sources:

ChatGPT ads sell worse than Google - for now, and based on someone else's data

In US traffic, paid Google results returned 2.45 dollars of revenue per session, compared with 0.89 dollars for ChatGPT ads. The transaction rate was 1.17 percent versus 0.37 percent.

Across 102 sites, 84 of 87 US advertisers earned less revenue per session from ChatGPT ads than from search. One agency spent over 1,500 dollars on 263 clicks and got a single form submission.

That is not a final verdict, though. The channel is young, and the authors measured only revenue per session - everything else, including any return multiple, is modelled, not measured.

What to do: before you move budget, compare both channels using your own costs and attributed revenue within the same attribution window. Someone else's CPC tells you nothing.

Sources:

Four topics you cannot use for audience targeting in Google Ads

Google Ads has clarified that advertisers cannot target users based on their interest in four categories of prohibited content: counterfeit goods, dangerous products and services, enabling dishonest behaviour, and inappropriate content.

In practice, this covers interest-based audiences and custom segments, and therefore audience signals in automated campaigns as well.

The change is mostly a formality - the policy now links the existing prohibited content rules to the personalised advertising rules - but during an account audit it gives you a concrete list to check.

What to do: review your audience and signal definitions and record the review date and the owner. Do not delete everything that looks sensitive - check whether the interest genuinely maps to a prohibited category.

Sources:

URL exclusions in AI Max only match whole words

Google has published documentation on URL exclusions in AI Max for Search campaigns. There are three methods: exact URL, page feed labels and a "URL contains" rule.

Text rules match only whole tokens, not parts of words. Underscores do not split tokens, and spaces act as an AND operator regardless of order.

Prerequisite: exclusions work only when text customisation and final URL expansion are both turned on. After migrating from dynamic search ads, your old rules become read-only.

What to do: test the rules against real URLs and export your old DSA rules before migrating, because afterwards you will not be able to edit them.

Sources:

One video in Performance Max means 12 percent more conversions, according to Google

In its Performance Max guidance, Google says campaigns with at least one video see a conversion lift of roughly 12 percent. You can build one in Asset Studio from product photos, your logo and short clips.

The second tip concerns lead forms: one or two extra questions, such as a postcode or the service needed, filter out leads you cannot serve. A reCAPTCHA blocks bots.

The third is about signals. Search themes and a list of your best customers uploaded through Customer Match do not limit reach; they simply give the model a better starting point.

What to do: add a video to a campaign that has none and compare conversions over the next two weeks - a new campaign goes through a two-week learning phase anyway.

Sources:

AI, agents and tools

Your customer uses free ChatGPT - and gets a fifth of what you get

Free ChatGPT runs about 1.5 to 2 search queries and checks roughly 20 pages. The paid version runs about 10 queries and reads close to 100 pages.

As a result, the answers can differ dramatically. The free tier has less material to work with, so it reaches for less obvious sources less often - and your content may be among them.

This is a problem for measuring AI visibility, because you do not know which tier you are measuring. A paid account shows a better picture than the average customer sees.

What to do: measure visibility on both tiers and report them separately. B2B customers are more likely to pay; e-commerce customers are more likely not to.

Sources:

Eight AI agents ran a company for a week and made zero sales

A company of eight agents ran on its own for seven days, selling a cheap guide. The agents wrote the guide in a matter of minutes, completed 721 tasks and sent 329 emails.

The result: zero sales. Most of the emails were delivered, a few replies came back, and the product was viewed a few dozen times.

The cause was not technical. The agents finished a task and then stopped until someone nudged them. On top of that, the agent acting as CEO invented a three-day wait before any follow-up, which froze sales.

What to do: use companies like this for work that already exists - content, documentation, testing, research - and give them an audience or a customer base, because without trust they will not sell anything.

Sources:

The Google Ads agent is read-only for now - and that is good news

The official Google Ads MCP server currently supports read-only access. An agent can analyse an account and report on it, but it cannot change budgets or bids on its own.

That matches real-world practice: the first task worth handing to an agent is reviewing search terms and hunting down wasted spend, not "optimising the account".

Keep in mind the budget limits that agents tend to forget. A campaign can spend up to twice its average daily budget in a single day and up to 30.4 times that budget in a month.

What to do: set thresholds for minimum clicks or spend, and ask for three alerts backed by evidence instead of twenty-five observations that require no action.

Sources:

Marketing, sales and content

Four hundred webinar signups is not a conversion problem, it is a room problem

In one programme sold through webinars, about 400 people register, 200 show up, 20 book a call and half of those buy.

The author of that funnel says there is no point optimising it. The issue is the quality of the room, not the order of the steps - with a good list, about 20 percent of attendees book a call, not 10.

The fix is scaling ads and scoring leads at the registration stage, not rewriting the sales page.

What to do: work out what share of signups could actually buy, and only then work on the conversion rate from attendance to call.

Sources:

Raise prices by 10 percent - the quickest brand strength test there is

If every customer leaves after a 10 percent price rise, you do not have a brand - you have a price. That is the test one investor and business owner uses to judge whether a brand is a real asset.

The M&A argument is similar: large funds bought well-known consumer companies for their brands, not for a novel business model.

For an agency, it is a question about its own offer. If clients choose on price alone, every price increase and every sales slowdown hurts twice as much.

What to do: find out what a client can say about you, apart from price, without opening your website. If the answer is nothing, your budget is going into price, not brand.

Sources: