> ## Content Index
> Fetch the complete content index at: https://bielawski.pro/llms.txt
> Use this file to discover other available public pages before exploring further.

# Press review 18.11.2026: CSP allowlists, a higher fee in France and a deal until 2031
- URL: https://bielawski.pro/en/press-review-18-11-2026-csp-allowlists-a-higher-fee-in-france-and-a-deal-until-2031/
- Published: 2026-11-18T06:10:31.000Z
- Updated: 2026-11-18T06:10:31.000Z
- Description: Malicious code reached 178 organisations through a single snippet, and only 15% of those scanned had a CSP policy that stopped it. From January, ads served in France and Italy carry an extra 3% fee.
- Author: Bartosz Bielawski
- Tags: Prasówka, #en

## SEO and AI search

### A CSP allowlist did not stop the malicious script, because the code came from a trusted domain

On 16 November an attacker added malicious JavaScript to customer projects in Optimizely Web Experimentation. The way in was a flaw in the account signup flow, and the code ran on every page that loaded the snippet, with no change to the page's own code.

Scott Helme of Report URI found the code in snippets belonging to 178 organisations, among them Zoom, BBC, CNN, Salesforce and GitLab. Of the 1611 snippet files analysed, 188 contained malicious code. As many as 68 organisations had no Content Security Policy at all.

Only 15% of the organisations scanned fully blocked the code from running, because they had a restrictive script-src. Adding the Optimizely domain to an allowlist does not help, because the malicious code was served from that very domain.

What to do: review script-src on every site that loads testing tools and tag managers. If the allowlist includes a provider's entire domain, it currently protects you from nothing.

**Sources:**

- Search Engine Journal, 2026-11-17: [Optimizely Says Attacker Used A Signup Flaw To Inject Malware](https://www.searchenginejournal.com/optimizely-signup-flaw-malware/597839/?ref=bielawski.pro)

### 88% of ChatGPT answers name no brand at all, and visibility still matters

In a panel of 41,961 ChatGPT conversations generated by 12,321 people, 6.9% showed purchase intent. The other 93.1% had nothing to do with buying.

The interesting part is what lies underneath. 87.9% of answers named no brand, product or business, and only 12.1% mentioned a brand. Images and visual design were the largest category of use.

Rand Fishkin points out that AI visibility is decided by Google's AI Overviews and AI Mode, not by ChatGPT alone, and asks whether tracking the chat is worth it without Bing rankings. Panel data tells you much more about what people ask than about how often they use the chat.

What to do: before you buy ChatGPT visibility monitoring, check whether your prompts trigger any brand mentions at all. Product questions do, but definition questions almost never do.

**Sources:**

- Search Engine Journal, 2026-11-17: [6.9% Of ChatGPT Conversations Showed Purchase Intent In U.S. Panel](https://www.searchenginejournal.com/chatgpt-purchase-intent-us-panel/597857/?ref=bielawski.pro)

### When everything is urgent, score the work out of 25

The GOAL-VP framework starts with two pass-or-fail tests. First: does the activity serve a business goal? Second: is it really about organic search? If it fails, you do not score it any further.

Only then do you calculate value as impact times certainty, each scored from 1 to 5\. The maximum is 25 points. Refreshing meta titles scores 15 out of 25, because impact is medium and certainty is high, and that is enough to make the list.

The author warns that a flood of "must-do" tactics lets non-SEOs start steering organic strategy. A value score lets you show the board which revenue is at risk if the work does not happen. LLMs.txt and markdown versions of an entire site are his examples of work with unproven returns.

What to do: take your current backlog and score every item the same way. Anything below 10 points does not make it into this quarter.

**Sources:**

- Search Engine Land, 2026-11-17: [How to prioritize SEO and GEO work when everything feels urgent](https://searchengineland.com/prioritize-seo-geo-work-urgent-502116?ref=bielawski.pro)

### A 72% AI visibility score can come from five branded questions

A 72% AI visibility score can describe a narrow set of branded questions, asked by one customer group, in one market, on one platform. That is not a measure of brand visibility; it is a measure of that prompt set.

Prompts are not chosen the way rank-tracker keywords are. They have to mirror buyer questions across the buying stages, including unbranded ones, and every model needs to be analysed separately, because each one retrieves and cites sources differently.

The author advises separating observed referral traffic from modelled impact in reporting. Low AI referral traffic does not prove low influence on purchase decisions, and Reddit's frequent presence in answers reveals an information need, not a channel to fill.

What to do: before you show the board a single AI visibility number, show the prompt list it was built from. Otherwise you are defending a number, not a brand.

**Sources:**

- Search Engine Journal, 2026-11-17: [AI Visibility Data: Avoid 4 Traps & Prioritize The Right Work](https://www.searchenginejournal.com/ai-visibility-data-traps-transperfectdigital-spa/597627/?ref=bielawski.pro)

## Google Ads and paid media

### An ad shown in France costs you 3% more, wherever your business is based

From 1 January 2027 Google adds a 3% Regulatory Operating Cost fee to ads served in France and Italy. What counts is where the ad is shown, not where the business sits, so a Polish or American advertiser pays it too.

In France the rate rises by a full percentage point, which means the fee itself is half again as high as before. In Italy it rises by half a point. The fee lifts your effective cost: a CPA of 40 euros becomes 41.20 euros, and a ROAS of 4 falls to about 3.88.

The catch is that the default location setting in Google Ads is „Presence or interest”, which Google labels as recommended. Google will not refund fees triggered by legitimate clicks and impressions.

What to do: treat the surcharge email as an instruction to audit your geo settings. Check whether campaigns pick up impressions in France and Italy through „interest” rather than physical presence.

**Sources:**

- Search Engine Journal, 2026-11-17: [Google Ads Surcharges Rise To 3% In France And Italy – Local US Advertisers Can Be Billed Too](https://www.searchenginejournal.com/google-ads-surcharges-rise-to-3-in-france-and-italy-local-us-advertisers-can-be-billed-too/596938/?ref=bielawski.pro)

## AI, tools and automation

### 25,000 pages in 45 minutes for 7 dollars, so check whether your team still does it by hand

Letaido is a marketing AI platform built by the Ahrefs team, with more than 70 integrations, including HubSpot, GA4, Notion, Stripe and Google Ads. Its Ahrefs connector offers around 525 actions across 17 tools, including data you cannot get through the public API.

The gap compared with the public MCP server is measurable: about 28 Site Explorer reports there versus 104 in the Letaido connector. An analysis of 25,000 pages took 45 minutes and cost 6-7 dollars in credits.

The 99-dollar plan covers unlimited users and 50 dollars of AI credits a month. Credits are used up mainly by long, multi-step builds on the most expensive model, while scripted automations cost pennies.

What to do: count the hours you spend each month refreshing data-driven articles and pulling Site Explorer reports. That is the first place a platform like this pays for itself.

**Sources:**

- Ahrefs Blog, 2026-11-17: [Letaido Review: 10+ Jobs Our Marketing Team Hands to It Every Week](https://ahrefs.com/blog/letaido-review/?ref=bielawski.pro)

### OpenAI's new model costs the same as Fable 5.2 and handles front-end work better

GPT-6.2 Astra is OpenAI's top publicly available model and a direct competitor to the Fable family. The price is identical: 8 dollars per million input tokens and 40 dollars per million output tokens.

The context window is a little over a million tokens, with a knowledge cutoff of 18 May 2026\. The model accepts text and images as input and returns text. In testing it clearly outperforms Fable 5.2 at building front-ends, sites and games.

After a series of tests on a 200-dollar-a-month plan, 64% of the allowance was left. The author notes that the model is fast and token-efficient even at a high reasoning effort.

What to do: if you build sites or prototypes with a model, run the same prompt on both and compare cost and time. At an identical price, the winner is the one that adds less of its own.

**Sources:**

- Bijan Bowen, 2026-11-18: [GPT-6.2 Astra Is INSANE – Does THIS Beat Claude Fable?](https://www.youtube.com/watch?v=jm%5Fh5nRZoCg&t=159s&ref=bielawski.pro)

## E-commerce

### Allegro and InPost signed a deal running to 2031, worth at least 8 billion zloty

Allegro and InPost signed a new multi-year framework agreement running to the end of 2031\. The minimum contract value is around 8 billion zloty over 2026-2031.

Because of the settlement, Allegro estimates its adjusted EBITDA for 2026 will be 270-280 million zloty higher. On the day it published its third-quarter results, the company reported 1.13 billion zloty of EBITDA, and its share price rose more than 4%.

The deal ends a dispute over delivery rates that had shaped shipping costs on Poland's largest marketplace for months. For sellers it means steadier rates and no risk of a carrier switch in the middle of the season.

What to do: if you are setting delivery costs in your 2027 pricing, base them on the new agreement, not on last year's negotiated rates.

**Sources:**

- Giełda Inwestycje Trading, 2026-11-17: [Inwestorzy UCIEKAJĄ z Azji! Allegro i InPost się w końcu dogadali.](https://www.youtube.com/watch?v=2Fou-Kwa2uo&t=1168s&ref=bielawski.pro)

## Marketing and content

### More than half of Americans fear AI more than they welcome it, so look at your own site

In a Pew survey, 52% of Americans are more concerned than excited about AI in daily life, while only 9% feel the opposite. The author argues that another product feature will not change that.

The site he built with AI for an instrument rental business would normally have needed a team of six and a budget of more than 60,000 dollars. The difference came from an expert interview and its transcript, not from a generic prompt.

Generic AI sites are recognisable by a single template: a big headline, a subheadline, rows of cards, a button. The advantage will lie not in how a site looks, but in how easily a visitor reaches their goal.

What to do: walk through your own site as a customer and count the steps between them and their goal. Changing the model will not fix that number.

**Sources:**

- Search Engine Land, 2026-11-17: [Vibe web marketing: How to manage your own AI team](https://searchengineland.com/vibe-web-marketing-ai-team-502169?ref=bielawski.pro)

### Hours and deliverables stop being the measure, because clients pay for the outcome

Newcomers using AI gain experience faster, because the feedback loop has become much shorter. Experience still wins, but for a different reason: it lets you challenge a model's recommendation because you have seen a similar situation and know how it ended.

The author says plainly that agencies pricing work by hours and deliverables are losing ground as content production becomes very cheap. The value of work comes from what it achieves, not from the time it takes to produce.

The key skill is now diagnosing the business problem before writing the prescription, alongside connecting AI to the site itself: API, MCP, CMS, repository. Delegating to agents is not the same as competence, because agents produce convincing but commercially useless output.

What to do: review what you price per item and per hour, and turn it into an outcome the client can measure.

**Sources:**

- Search Engine Land, 2026-11-17: [When everyone can do SEO, what makes an expert?](https://searchengineland.com/everyone-can-do-seo-what-makes-expert-502136?ref=bielawski.pro)

### A five-second test shows whether your site is selling the right service

Google returns the pages it considers most relevant to a query. A site that mostly describes one service will be found for that service, not for the new one you want to start selling.

The test is simple: show the page to an outsider for five seconds and ask what the company sells. If they name the old service, your site is selling the wrong thing, even though that service is not in your offer.

After one client's site was rebuilt, some keywords moved from the third page of Google into the top five positions. The author points out that nobody can guarantee a number one ranking, and Google says the same.

What to do: run the five-second test on your most important service page and compare the answer with what you want to sell next year.

**Sources:**

- Missoula SEO Geek, 2026-11-17: [Their Website Was Selling the Wrong Thing](https://www.youtube.com/watch?v=HxGxchMsmMQ&t=30s&ref=bielawski.pro)