In short and in plain language
Many firms let AI write a first draft of their website texts. AI means artificial intelligence. It is a program that suggests text.
Google says: what counts is whether a text helps the reader. It does not matter who wrote it.
But many pages that all say the same and help no one can be spam. Spam here means pages made only for search. Google acts against such pages.
Check each text before it goes online. Are the figures right? Are the prices and dates right? The title and the text for each image count too.
Add what only you know. Show photos of your own jobs. Explain how you work. We can help you with this.
Service descriptions, guide articles, page titles and meta descriptions are now written with AI assistance in many businesses. The same question follows almost every time: does Google penalise this kind of copy? The answer is in Google’s own documentation, and it is more sober than many rumours. What counts is not whether a person or a program wrote the first draft, but whether the page offers visitors added value. Anyone who generates many pages that help nobody, on the other hand, may violate the spam policy on scaled content abuse. This article sets out what Google actually says about AI-written copy, where the line runs, where AI genuinely helps in the editorial workflow and what a review process for service-business websites looks like, one that covers facts, figures and metadata and treats first-hand experience as the part no draft can supply.
Key points at a glance
- Google judges content by quality, not origin: according to Google Search Central, using AI gives content no special advantage, but appropriate use does not violate the guidelines either.
- Many pages without added value can count as spam, whether people or AI produce them. After the March 2024 update, Google reported 45 percent (Google) less low-quality, unoriginal content in its results.
- AI helps the editorial workflow with outlines, first drafts and variants: 35 percent (Destatis) of companies using AI generate natural language with it. Prices, deadlines and credentials come from your own records.
- Before publishing, check every figure and every legal statement at the source. Invented qualifications, memberships or awards can be misleading under Section 5(2) no. 3 UWG, the German unfair competition act.
- Google explicitly extends the review to metadata: title elements, meta descriptions, structured data and alt text can appear in search results and therefore need to be read and validated one by one.
- First-hand experience is the difference: add workflows, project photos and insights from your own jobs, name authors openly and record the sign-off with name and date.
What Google actually says about AI-written copy
Google set out its basic position back in February 2023 (Google Search Central) in a post on AI-generated content, and nothing fundamental has changed since. Its ranking systems aim to reward original, high-quality content that demonstrates experience, expertise, authoritativeness and trustworthiness, known as E-E-A-T. How a piece of content was produced is not the focus. In the FAQs accompanying that post, Google writes that using AI gives content no special advantage: if a text is useful, helpful and original, it can perform well, and if it is not, it may not. Conversely, appropriate use of AI or automation does not violate the guidelines. What is not allowed is using it with the primary purpose of manipulating rankings in search results.
For website owners there is now dedicated Google Search guidance on using AI-generated content on their own site, which Google keeps up to date. It opens with a point worth remembering: generative AI can be particularly useful when researching a topic and adding structure to original content. The very next sentence draws the line, and that line is the heart of the whole question.
However, using generative AI tools or other similar tools to generate many pages without adding value for users may violate Google’s spam policy on scaled content abuse.
The guidance also points to the Search Quality Rater Guidelines. There, sections 4.6.5 and 4.6.6 (Google Search Central) describe how raters evaluate scaled content abuse and main content created with little effort, little originality and little added value. These raters do not rank pages. According to Google, their ratings do not directly influence ranking but are used to measure how well the search systems perform. A figure for 2023 shows how large this measurement effort is: Google says it ran over 700,000 experiments (Google) that resulted in more than 4,000 improvements (Google) to Search. For a business with a website, this means a trick that still slips through today is no reliable foundation. How content appears as a source in AI answers in Search is a separate topic, covered in our article on AI Overviews and visibility.
The line: scale without added value
In March 2024, Google introduced three new spam policies (Google Search Central), among them the policy against scaled content abuse. The definition is deliberately independent of the tool: it covers many pages generated primarily to manipulate search rankings rather than to help users. Typically this means large amounts of unoriginal content that provides little or no value, regardless of whether a program, a team of writers or both produced it. The older policy against automatically generated content was thereby explicitly extended: producing content at scale to manipulate rankings is abusive, whether automation or people are involved.
Google’s own figures on this update show how seriously it takes the issue. At launch, Google expected the combination of the update and earlier efforts to reduce low-quality, unoriginal content in search results by 40 percent (Google). After completing the rollout in April 2024, the company reported a reduction of 45 percent (Google). Such values cannot be verified from outside, but they show the direction: pages that exist only to cover one more search query are losing ground. The spam policies give examples that occur on small websites more often than one might think.
- Using generative AI tools to produce many pages without adding value for users
- Taking feeds, search results or other people’s content and reshaping it through synonymising, translation or other obfuscation
- Stitching together content from different web pages without adding value of your own
- Creating multiple sites to hide the scale at which content is produced
- Filling many pages with search terms whose text makes little or no sense to a reader
A typical case for service businesses
For your own website, this leads to a simple test before every new page: would a customer miss it if it did not exist? Does it answer a question that the existing pages leave open? If neither can be answered with yes, a better prompt will not help either. Existing pages of this kind belong in a content audit, where they are merged, revised or removed. Warning signs include groups of pages with almost identical text, guides without a recognisable author and metadata that is the same across many pages.
Where AI helps in the editorial workflow
AI is spreading quickly in German companies, including for writing. According to the ICT survey of the Federal Statistical Office, 26 percent (Destatis) of companies with ten or more employees used artificial intelligence technologies in 2025. Of these companies, 35 percent (Destatis) used AI to generate natural language including program code; in the size class with 10 to 49 employees the figure was 32 percent (Destatis). 52 percent (Destatis) of the companies using AI had images, videos or audio generated. A representative survey of 603 companies with 20 or more employees on behalf of the digital association Bitkom found in September 2026 that 57 percent (Bitkom) use AI. Among the areas of use, marketing and communication follows customer contact with 54 percent (Bitkom). The two surveys differ because their populations and survey years differ, but they point the same way.
So where does a language model actually help on a service-business website? Google names researching and structuring original content. In practice, there are further tasks where a draft saves time without having to supply facts: an outline for a new service page, a first draft from bullet points, several variants for a page title, shortening a long paragraph or collecting the questions customers typically ask about a service. In all these cases the result is a starting point that someone who knows the business continues to work on. The same applies to ad copy, which Google additionally reviews before it is served; our article on Google Ads for Black Week shows how sale ads are prepared.
| Task | Where the draft helps | What a person checks |
|---|---|---|
| Outline of a service page | Order of questions, subheadings | Do scope and process match the actual offer? |
| First draft from bullet points | Sentence structure, readability, length | Facts, technical terms, the business’s tone |
| Page title and meta description | Several variants to choose from | Unique per page, no keyword stuffing, matching the content |
| Alt text for images | Suggested wording | Does the image really show what the text says? |
| FAQ draft | Collecting typical customer questions | Checking answers against your own rules, deadlines and prices |
| Figures and legal statements | Flagging where information is missing | Every figure with a source, every rule checked against the statute |
Another figure from the Bitkom survey is revealing: among companies already using AI, only 26 percent (Bitkom) complain about poor quality of results. That sounds reassuring, but for website copy it is more of a warning sign. A fluently written paragraph seems credible even when a deadline, a price or a certification in it is wrong. Precisely because drafts sound good, a fixed review step is needed, one that depends not on impressions but on a checklist.
First review step: facts and figures
Google explains the need for review in technical and unusually clear terms. Generative models do not retrieve facts; they predict a likely sequence of words based on their training data. As a result, their outputs may contain inaccuracies, also known as hallucinations. The guidance therefore calls it critical to manually fact-check and review all AI-generated content for accuracy and trustworthiness before publishing. For a service-business website, this translates into a short, binding checklist that is worked through before every sign-off.
- Every figure has a source that is named in the text, and the figure still appears at that source today.
- Prices, terms and deadlines come from your own current records, not from the draft.
- Legal statements are checked against the statute, with section and subsection, not merely phrased plausibly.
- Qualifications, certificates, memberships and awards appear in the text only if the business holds proof of them.
- Names of staff, places and projects are spelled correctly and may be published.
- Every link leads to the intended page, not to a homepage or an error page.
The fourth point is more than diligence. Language models like to add whatever seems to fit a service description: a master craftsman’s certificate, a guild membership, a certification, a founding year. Under German competition law, anyone who engages in a misleading commercial practice that is likely to cause consumers or other market participants to take a decision they would not otherwise have taken acts unfairly (Section 5(1) UWG). The law explicitly names untrue or deceptive statements about the trader’s qualifications, status, approval, memberships or awards (Section 5(2) no. 3 UWG). Responsibility for the text stays with whoever publishes it, whatever tool suggested the sentence. This section explains the legal position for work on a website and does not replace legal advice in individual cases.
In practice, one simple rule works well: anything a customer could later check or claim is not taken from the draft but inserted from the source. This includes prices, scope of service, response times, warranty and dates. The draft may mark the place where such information belongs; the information itself comes from the quote, the contract or the price list. How to structure service copy that shows benefits rather than features is covered in our article Website copy that sells.
Second review step: titles, descriptions, data and alt text
One sentence in the guidance is often overlooked, although it has the most consequences for day-to-day work. According to Google, the review also applies to metadata like title elements, meta descriptions, structured data and alternate texts for images, because they can appear in search results. These fields in particular are often generated in a single run, for many pages at once, and then read by nobody. Each of the four fields has its own rules.
Title
Every page gets its own descriptive title. Google advises against vague labels such as “Home”, boilerplate repeated on every page and stuffed keywords.
Meta description
It summarises the page in a short, informative overview. According to Google, descriptions made of long strings of search terms are less likely to be shown as a snippet.
Structured data
The markup describes only what visitors see on the page. Hidden or misleading information can lead to a manual action.
Alt text
It describes what the image shows, using fitting terms but without keyword stuffing, and at the same time serves people who cannot see images.
Page titles deserve particular care. Google generates the title link in search results itself, but according to its own statement from 2021 it uses the content of the title element more than 80 percent (Google Search Central 2021) of the time. A generic title from a mass run is therefore very likely the first thing a searcher sees of the page. Good titles name the service and what sets it apart, such as location, audience or a special feature, and do not repeat the same building block on every page.
For meta descriptions, Google primarily builds snippets from the page content and uses the description when it summarises the page more accurately than the text itself. For large, database-driven sites where writing descriptions by hand is not realistic, Google even recommends generating them automatically and asks for human-readable, varied text. A service-business website with a manageable number of pages is not a database: here, a hand-checked description per page is the norm, and the effort stays small.
Structured data is where an error is least noticeable, because no visitor sees it. Google requires the markup to be representative of the main content of the page and to mark up only content that is visible to readers. An AI draft that writes opening hours, reviews or prices into the JSON-LD that do not appear on the page risks a manual action. The consequence is that the page can no longer appear as a rich result; according to Google, such an action does not affect ranking in web search. After every change the markup belongs in a validator, as the AI guidance also requires. Our article on rich results and schema markup shows how to use structured data properly.
With alt text, search and accessibility are closely linked. Google calls it the most important attribute for providing metadata for an image and warns against stuffing alt attributes with keywords. An automatically generated alt text often only describes what an image recognition model assumes. Only the business knows whether the photo really shows its own workshop, the finished project or a stock image. Anyone using AI-generated images in an online shop should also know that for ecommerce sites, Google Merchant Center requires AI-generated images to carry the IPTC DigitalSourceType TrainedAlgorithmicMedia metadata. More on formats, file sizes and image descriptions is in our article Optimise website images.
First-hand experience as the difference
Google sums up its quality signals as E-E-A-T: experience, expertise, authoritativeness and trustworthiness. Of these aspects, Google says trust is the most important, and the others contribute to it. E-E-A-T is not a single ranking factor but a mix of signals. For self-assessment, Google suggests questions like these: does the content clearly demonstrate first-hand expertise, for example because someone has actually used a product or service or visited a place? Does the page provide substantial new information, or does it merely summarise what others have already written? And does it offer substantial value compared with other pages in search results?
This is exactly the part a language model cannot supply. It knows no building site, no customer and no job of your business. An AI draft for a “bathroom renovation” service describes what a bathroom renovation is in general. A plumbing firm’s page only becomes helpful through what only this firm knows: how coordination with the other trades works, which questions come up regularly at the first appointment, how you can still tell a well-planned bathroom years later and which surprises are typical in older buildings in the region.
The question for every draft
Sources of first-hand experience exist in almost every business anyway: the questions customers ask on the phone or by email, the workflow from first contact to acceptance, photos from your own projects with customers’ consent and the reasons why the business turns down certain jobs or solves them differently from the norm. If you collect these points as bullet points before writing and give them to the draft, you get a better starting text and know exactly which paragraphs still need adding afterwards. Our overview of websites for service providers shows how such pages are structured for a sector.
Experience also means that it is clear who is writing. Google recommends bylines where readers expect them and explicitly warns against invented authors: AI-generated portrait photos, made-up names or false credentials meant to make content look as though it was written by human experts are treated by Google as deception. A guide article showing the name and role of the person who is responsible for its content, by contrast, is a simple and honest trust signal.
Disclosure: how a text was created
Does every text need a note saying that AI was involved? For Google Search the answer is: not necessarily, but where it helps readers. Google advises making the “how” of content visible when someone might wonder how it was created. According to the guidance, anyone generating content automatically should consider adding information on how it was created in a way that makes sense for their audience, for example background on how automation was used or image metadata. A byline naming the AI as author, on the other hand, is not necessarily the right way to tell readers that AI was involved, in Google’s view.
For a typical service-business website this suggests a pragmatic line. Service copy that an employee drafts with AI assistance and takes professional responsibility for usually needs no separate note from the perspective of Google Search. Content that is created largely automatically, such as summaries, machine-translated versions or chatbot answers, is a different case. Whether a statutory labelling obligation also applies there is a separate legal question, covered in our article on labelling AI chatbots. For ranking, it is not the yardstick.
A review process for service-business websites
The requirements can be turned into a process that works in small businesses too. It needs no special software, just clear responsibilities and a fixed order. What matters is that every step produces a visible result: a bullet-point list, a marked-up draft, a ticked checklist, a sign-off with a name.
- Briefing with your own facts: note the service, audience, workflow, service area and typical customer questions as bullet points before any draft is created.
- AI draft: have an outline and first draft generated, explicitly without invented figures, names or credentials; mark missing information as placeholders.
- Fact check: check every figure, every deadline and every legal statement at the source and name the source in the text.
- Add experience: insert workflows, project photos and insights from your own jobs, shorten or delete generic paragraphs.
- Check metadata: read the title element, meta description, structured data and alt text one by one and validate the markup.
- Sign-off: the person responsible for the content reads the full text and records the name and date of the sign-off.
- Review date: set a date on which prices, deadlines and statements are checked again.
Signing off with name and date looks bureaucratic but has a practical benefit. If a customer asks about a statement months later, it is clear who checked it and on what basis. On websites maintained by several people, the sign-off also prevents an unchecked draft from going live by mistake. Whether Google has already taken a manual action against a website is shown in the Manual Actions report in Search Console; a look there belongs in every review of existing content.
How to plan new content from the outset instead of tidying it up afterwards is described in our article on content marketing strategy for SMBs, and the craft of writing itself in How to write SEO content. Our content marketing combines both: we plan topics with you, write service pages and guides, check figures, titles and descriptions and build in your project experience so that it counts for visitors and for search engine optimisation.
Related Articles
Content Marketing Strategy for SMBs: the Plan
How small and medium businesses build organic visibility over time with topic clusters, an editorial calendar, SEO texts and honest measurement.
Checking SEO or GEO offers: questions before hiring
How to check SEO and GEO offers: Google's questions for providers plus checks on reporting, access, contract term and AI texts, without ranking promises.
Content audit: cleaning up old website content
Old pages, duplicate topics, outdated prices: how a content audit sorts an existing inventory — with four exits, fixed criteria at the entrance and redirects.