SEO with AI: what helps and what hurts you

What Google actually says about AI content

The official line has been unchanged since early 2023 and is clearer than many discussions suggest: Google judges the quality of content, not how it was produced. Using AI is therefore not prohibited. A text is not ranked lower because a model helped write it, and not higher because a human typed every word.

Google draws the line at purpose. The spam policies describe scaled content abuse: generating many pages primarily to manipulate rankings while providing little value is a violation, no matter how the pages were created. The first example Google gives is explicitly generative AI used to produce many pages without adding value for users. Also listed: automatically rephrased or translated third-party content and text stitched together from other pages.

Google also prohibits site reputation abuse: publishing third-party content on a well-regarded domain so it can ride on that domain's reputation. For company sites this rarely comes up. It becomes relevant when someone offers you paid guest posts that have nothing to do with your business.

On disclosure, Google is pragmatic. It recommends bylines wherever readers would ask who wrote this, and a note on AI use wherever they would ask how it was created. Listing the AI itself as the author is something Google explicitly calls probably not the best approach.

What AI genuinely does well in SEO

The benefit lies almost entirely upstream of the text. AI is strong at pulling the recurring questions out of a handful of customer conversations and turning them into a topic list. It is fast at checking existing pages for missing subheadings, duplicate topics and unanswered questions.

Also worth it: generating structured data, drafting meta descriptions for thirty pages in one pass, preparing translations, proposing internal links. These are mechanical jobs with clear rules, and that is exactly where the time saving is real.

What remains: somebody has to know the subject and check the output. An invented figure in a meta description goes live faster than anyone notices it, and then sits in the search results for months. The overview below shows how we rate the typical uses.

Why machine-written text alone does not rank

Google does not judge whether a text came from a human, it judges whether it is useful to someone. That sounds relaxed and is not, because mass-produced text almost never meets the condition. It summarises what is already online twenty times over and delivers not one sentence that did not exist before.

The difference is rarely style, it is substance. A number from your own operation, a case that went wrong, a concrete price range, a rule you worked out yourself. None of that is in the model, because it comes out of your work and not off the web.

The workable route is therefore not to keep AI away from writing, but to reverse the order: you supply what only you know, and the machine puts it into shape. Not the other way round.

A workflow for AI-assisted articles

This is the workflow we use to produce specialist articles like this one. It is not faster than pure generation, but it is clearly faster than writing without support, and at the end there is a text someone can stand behind.

1. Fix the question. Which concrete search question should the article answer, and who asks it? One article, one main question.
2. Collect knowledge before the AI writes. Notes, numbers from your own projects, typical customer objections, mistakes you have seen. A recorded twenty-minute conversation with the expert is often enough.
3. Build the brief with AI. Outline, questions the article must answer and gaps in competing pages. This is where the model is strong.
4. Draft from your material. The AI writes from the collected notes, not from its general knowledge. Anything that did not come from the material gets flagged.
5. Check sources. Every statement about rules, laws or platforms is checked against the primary source, meaning the provider's documentation or the legal text. No number without an origin.
6. Expert sign-off. The expert reads the text and cuts whatever they would not say that way. This is the step skipped most often, and the most important one.
7. Technical work and markup. Title, description, structured data, internal links, author and date. Here the AI can take on more again.
8. Maintain. After a few months, check in Search Console which queries the article appears for, and update it when rules change.

The test for every paragraph: could a competitor put it on their site word for word? If so, something of yours is still missing.

The real shift: answers without a click

More important than the question of who writes the text is a different one right now: more and more searches end with an answer inside the search engine, without anyone opening a page. Google shows summarised answers above the results, and a growing share of users reads only those.

Plenty of figures circulate on the scale of this, mostly from vendors who sell SEO tools. Shares well above forty percent of searches carrying a summarised answer get quoted, along with noticeable declines in clicks on the top positions. Treat those figures as a direction rather than a promise, they come neither from Google nor from your own data.

What follows is the same regardless of the exact number: first place no longer guarantees a visit. This hits general informational queries far harder than concrete buying intent, and for B2B services the valuable part still sits where someone is looking for a provider rather than a definition.

Visible in AI answers: what makes a citation more likely

When a machine assembles the answer, it also decides whom to name. That applies to Google's AI Overviews and AI Mode just as much as to ChatGPT search and Perplexity. Google is remarkably sober about it: there are no additional requirements and no special optimisation for AI Overviews and AI Mode. A page must be indexed and eligible to be shown with a snippet, nothing more. Special AI text files or extra schema markup are not needed either, according to Google.

Google also describes how both features can split a question into several related searches, the so-called query fan-out. The answer then draws on pages that answer the sub-questions well. From this follows what, in our experience and in widespread industry opinion, helps. These are not confirmed ranking factors, they are good practice:

• Direct answers. The key statement sits in the first sentences under the heading, not in the third paragraph.
• Clear structure. Headings that stand on their own as a question or a statement, plus lists and tables for comparisons and processes. Blocks like these are easier to quote than running prose.
• Original data. A number that only exists on your site is a reason to cite you specifically. A summary of what everybody writes is not.
• Consistent company details. Name, address and description read identically everywhere: on the site, in the Google Business Profile, in directories and in your profiles. Contradictions prevent unambiguous attribution.
• sameAs in the structured data. The Organization markup points to your official profiles, such as LinkedIn or the company register entry. That makes clear that all the entries refer to the same company.
• Do not forget Bing. ChatGPT searches the web with its own crawler and external search providers. In the industry Bing is considered an important source for this. OpenAI does not document it in detail, but registering with Bing Webmaster Tools takes ten minutes.

And the dullest point stays the most effective: content someone links to or mentions voluntarily, because it contains something they cannot get elsewhere. That was true before AI and has become more important since.

Search crawlers and training crawlers: what blocking means

Many companies block AI crawlers wholesale in robots.txt and later wonder why they do not show up in ChatGPT. The reason: there are two kinds of crawlers. One kind fetches pages so they can be linked in search results and answers. The other collects material for training models. Lumping both together costs visibility without anyone noticing.

For most B2B companies the calculation is simple: allow search crawlers, because you want to be found. Training crawlers are a matter of principle with no direct effect on your visibility in search. If you sell expertise that should not flow into models for free, block them. If you want to be known as widely as possible, allow them.

Check not only robots.txt but also your CDN. Some providers now block AI crawlers by default. Check there whether search crawlers such as OAI-SearchBot are affected too.

E-E-A-T: signals you can set yourself

E-E-A-T stands for experience, expertise, authoritativeness and trustworthiness. Google stresses that E-E-A-T itself is not a single ranking factor, it describes what the systems try to recognise through many signals. The most important of them is trust. For AI-assisted content that means: the machine cannot have experience. That has to come visibly from you.

• A real author. Name, role and an author page that shows where the knowledge comes from. No faceless editorial team.
• First-hand experience. Phrases like 'with our clients we see' backed by a concrete example. For us, for instance: more than 2,000 meetings booked since 2023, which tell us which questions decision-makers really ask.
• Your own data and evidence. Screenshots, measurements, before-and-after comparisons, with a note on how they were produced.
• Sources. Links to primary sources wherever you describe rules or platforms, as at the end of this article.
• Date and upkeep. A visible and machine-readable publication and modification date, and updates whenever something changes.

Much of this is a matter of page structure rather than text. How a page should be built so that people and machines understand it is covered in modern UX/UI design.

How to spot this in your own numbers

The view that shows it is free in Google Search Console. Compare impressions and clicks separately over the last twelve months. If impressions hold or rise while clicks fall, you are looking at exactly this effect and not at a ranking problem.

Confusing the two means optimising the wrong thing. A page still sitting in position three with fewer clicks does not need new backlinks, it needs an honest decision about whether that topic can still bring you visitors at all.

The topics that stay stable are the ones with buying intent and a local or specialist angle. The topics that break away first are definitions, listicles and what-is questions. That is a workable basis for deciding where your time still goes.

What you can and cannot measure about visibility in AI answers:

• Measurable: clicks and impressions from AI Overviews and AI Mode. Google counts them in the Search Console performance report under the Web search type, together with all other results.
• Measurable: visits coming from chatgpt.com, perplexity.ai or other AI services, as referral sources in your web analytics.
• Not cleanly measurable: how often you are named in an AI answer without anyone clicking. Tools that estimate this ask sample questions and count mentions. That is a sample, not a statistic, and the answers change from one request to the next.
• Simplest indicator: ask new leads in the first call how they found you. In our experience the answer 'ChatGPT recommended you' now comes up.

What to leave alone

Three things keep showing up right now and do more harm than good. First: generating a hundred pages in one go to cover topics. That produces a pile of thin pages competing with each other, and weakens the few good pages along with them.

Second: having the machine rewrite existing good pages so they sound more modern. That regularly swaps out precisely the concrete passages that made the text useful in the first place.

Third: tools promising to place your brand inside AI answers. The part of that which works is clean markup on your own site, and that carries no surcharge. The rest is sales.

Honest limits

Nobody outside Google, OpenAI and Perplexity knows exactly by which criteria sources are chosen for an answer. Anything beyond the official documentation, including in this article, is experience and observation, not certainty. Whoever promises you guaranteed mentions in ChatGPT is promising something they do not control.

On top of that, the products change faster than guidelines. What is true in September 2026 may look different six months later. Only the foundation stays stable: a technically clean site that loads fast and can be crawled, content with real substance and a company that is described the same way everywhere online.

How we handle it

On this site AI runs along in the preparation and in the technical work, not as a source of text. The numbers and cases come out of our work, the structure, the markup and the translations are produced with support. Every article carries an author and a machine-readable date, every page its own description.

If you need a site built so that search engines and AI services understand it and visitors still get in touch, here is how we do it: websites by BGS.

If you use AI motifs in advertising, a labelling duty has applied since August 2026. That concerns ads rather than your copy, but it is the point where the two topics touch: labelling AI content.

And when the visitors are there and nothing happens anyway, the search engine is rarely the reason. We wrote something about that here: conversion rate.

Frequently asked questions

Sources

1. Google Search Central: Google Search's guidance about AI-generated content
2. Google: Spam policies for Google web search
3. Google: Creating helpful, reliable, people-first content
4. Google: AI features and your website
5. Google: Overview of common crawlers, including Google-Extended
6. OpenAI: Overview of OpenAI crawlers