What Europe’s draft AI scraping guidelines mean for businesses

The European Data Protection Board adopted draft AI scraping guidelines for consultation on 7 July 2026. They explain how existing GDPR duties apply when businesses collect web data to develop generative AI.

As of 16 October 2026, Guidelines 03/2026 remain open for consultation, with comments due by 30 October 2026. That is the feedback deadline. Businesses already processing personal data must assess their activities under the existing GDPR framework. The practical recommendations below explain how to turn that assessment into collection and purchasing decisions.

Who the EDPB AI scraping guidelines affect

The draft focuses on private organisations using external personal data for generative-AI training and fine-tuning. It covers in-house scraping, commissioned collection and reuse of bought scraped datasets. Processing an organisation’s own personal data and broker-only activities are outside its stated focus, without being exempt from GDPR.

Consider three hypothetical projects:

Business project	Decision to make
A retailer collects product codes, prices and availability for a competitor dashboard	Check whether any part of the workflow processes personal data. Define the monitoring purpose before deciding whether to add other content.
The retailer adds customer reviews to fine-tune a shopping assistant	Assess why identifiable reviews are needed and what the intended model will do with them.
A software company buys scraped forum posts to fine-tune a support model	Review the collection history and proposed reuse before accepting the dataset into training.

Personal data includes information that identifies someone indirectly, as well as obvious names or contact details. Review text, photographs or combinations of fields can matter. Exporting only product columns does not establish that no personal data was processed during page fetching, rendering or intermediate storage.

Assess retrieval-only AI applications separately: the draft’s training focus does not resolve every AI architecture. Website terms, copyright and other applicable rules also need their own assessment.

Turn the purpose into a collection specification

The GDPR’s data minimisation principle already requires personal data to be relevant and limited to what is necessary. Its purpose limitation principle also matters when an existing dataset is proposed for a different use, such as moving from monitoring to model training.

Imagine a business developing an assistant that explains washing-machine specifications. Its initial brief says, “Collect every page from these appliance websites.” That could include customer reviews, reviewer biographies and support discussions alongside technical information.

A useful brief would identify approved product-page patterns, required fields and excluded sections. For this hypothetical task, the choices might look like this:

Proposed collection	Exclude unless separately justified
Model number, dimensions and energy rating	Reviewer names and profile links
Relevant manufacturer descriptions	Unrelated customer comments and biographies
Price, currency and availability where the task needs them	Personal contact details and account information
Source URL and collection time for checking records	Whole-page text retained simply because it is convenient

These are example design choices, rather than a regulator-prescribed schema. Selecting fields can reduce unnecessary retained data, but source selection and the processing that happens before export still need review.

With Web Scraper, field selectors define what a sitemap extracts. Inspect representative output, including long descriptions and unusual page layouts. A column named product_description might contain customer text after a website redesign.

Keep a record of the extraction rules associated with each reviewed batch in your own dataset documentation. This gives the team a way to investigate unexpected content before it reaches training.

Legitimate interests needs a specific explanation

The EDPB’s earlier Opinion 28/2024 explains three cumulative requirements for relying on legitimate interests: a legitimate interest, necessity and a balancing exercise against individuals’ interests and rights. Necessity includes considering equally effective, less intrusive ways to achieve the objective.

For the appliance assistant, those questions become concrete:

1. What present business objective does the training serve?
2. Could specifications and manuals achieve that objective just as effectively?
3. If customer experiences add necessary value, why does the dataset need identifiable comments, and what could happen to the people concerned?

The answers should change the collection brief. A model learning technical vocabulary has a different data requirement from one intended to summarise individual customer experiences. Removing authors’ names may still leave identifiable details inside the comments.

The EDPB’s existing AI opinion also makes reasonable expectations contextual. Source, relationships, publication circumstances and intended processing matter; providing a privacy notice does not by itself establish that people should expect a particular AI use.

The scraping draft recommends excluding websites that clearly oppose collection through technical measures such as authentication, robots.txt, ai.txt or CAPTCHA intended to prevent robot access. It also states that an absent robots.txt file does not amount to GDPR consent.

For businesses, this makes source restrictions part of the approval decision. Record the relevant signals and put disputed sources back through review before investing further in collection.

Make supplier evidence part of the purchase

Under the EDPB’s final controller and processor guidance, roles depend on actual decisions about purposes and essential means. A contractor following documented instructions may be a processor; independent collection and reuse may involve separate controllers; shared decisions can create joint control. Contract labels alone do not settle those roles.

For a commissioned dataset, agree who decides the sources, personal-data categories, retention and permitted uses. For an existing corpus, request an inspectable sample and collection record. Useful questions include:

• Which sources and collection periods does the file cover?
• What personal data does it contain, and what was excluded?
• What evidence supports the original collection and the proposed reuse?
• How are corrections, objections and justified removal requests passed between the parties?
• What happens to stored copies when the service or agreement ends?

These questions expose missing information before a file becomes a training dependency. A lower-priced dataset with little source history may require more assessment work than one with a usable collection record. A supplier’s general assurance cannot answer whether the data suits your particular project.

Give transparency a workable process

Collecting personal data indirectly brings information duties. These include explaining the purpose, legal basis, sources, retention and relevant rights.

An exception may apply where individual notification is impossible or requires disproportionate effort. The EDPB-endorsed transparency guidance requires a narrow, documented assessment. Where the exception applies, the controller must still make the required information publicly available and take appropriate protective measures.

Test the handover with a hypothetical enquiry. Could support identify the dataset owner, explain where the data came from and route the request to someone able to act? If the answer depends on finding an old supplier email, improve the process before the dataset becomes routine.

Treat sensitive content as a separate decision

Health information, political opinions and other Article 9 categories have additional protection. The EDPB explains that processing such data requires both an Article 6 lawful basis and an Article 9(2) exception.

The draft discusses whether search-engine case law may be relevant to incidental, residual collection despite prevention measures. The EDPB expressly rejects a general exemption and requires a case-specific assessment.

In the appliance example, a review might identify a customer and reveal a medical condition. Removing email addresses would not detect all such content. Source choice, sample review and escalation criteria matter alongside filters.

If unexpected sensitive material appears, hold the affected batch out of training while the responsible team assesses it. The collection process needs a response to that discovery, rather than a general promise to clean the dataset later.

Follow changes through to the model team

Erasure and objection rights are subject to conditions and exceptions. Where processing relies on legitimate interests, continuing after an objection may require demonstrating compelling overriding grounds. CNIL’s AI guidance also distinguishes requests affecting training datasets from those affecting a model. Deleting a dataset row does not necessarily remove information already learned by a model.

As a practical test, choose one source and ask which stored batches contain it, which dataset versions used those batches and which model versions received them. Keep proportionate source history that helps answer those questions, with a named owner for downstream action.

For recurring collection, also decide how an approved exclusion or correction affects future runs. Otherwise, the next refresh could reintroduce material the team has already removed.

Where Web Scraper fits

Build and test the extraction sitemap in the Web Scraper browser extension, inspect the resulting records, then import or sync it to Cloud. Configure scheduled Cloud runs once the tested extraction meets the specification.

Web Scraper provides the extraction and delivery layer. Source approval, privacy assessment, personal-data filtering, dataset governance and model controls remain part of the customer’s wider workflow.

Before accepting the next training batch, have its owner resolve any missing source evidence and confirm who handles downstream changes. For the collection stage, explore Web Scraper’s workflow for AI training and evaluation data.