How to Use AI for SEO: What Actually Works and What Wastes Time

TL;DR: AI works on preparation and fails at publication. Ahrefs studied 331,000 pages in June 2026 and found no penalty for AI authorship, but heavy AI pages earned two to three times fewer impressions and were indexed 40.35% of the time against 49.28% for lightly assisted pages. Google's July 2026 documentation lists the AI tactics you can ignore, including llms.txt. The waste sits in two places: publishing machine drafts at volume, and buying tactics Google has already said do nothing.

What using AI for SEO actually means

Using AI for SEO means applying language models to the research, analysis, drafting and reporting inside a search programme, rather than to the act of publishing pages. Where you draw that line decides whether the work compounds or costs you.

Google's position has not moved. Its guidance says generative AI is particularly useful when researching a topic and adding structure to original content, then warns that using it to generate many pages without adding value may violate the scaled content abuse policy.

That policy is worth reading closely. Google defines scaled content abuse as many pages generated primarily to manipulate rankings rather than help users, and describes it as large amounts of unoriginal content providing little value, no matter how it is created. The rule is method-neutral. Human writers producing thin pages at volume fall under exactly the same policy.

If the surrounding vocabulary is unfamiliar, our field guide to GEO, AEO, AIO and LLMO covers what each label actually refers to.

Why AI content and bad content overlap

Quality matters. How it was produced matters much less. The two keep showing up as the same thing because of the effort behind them, not the tool.

Here is what most people do. They give a model the most basic input they can get away with, publish whatever comes back, and expect citations. Think about that logically. You took information out of an LLM, added nothing, published it, and now you want that same LLM to cite you as a source for information it already had. There is nothing in that loop for the model to reward.

Nobody is reading thin AI pages, and that includes the language models. The page holds no observation the model lacked, no data it cannot generate itself, and no experience it has no access to.

What works is the opposite. Answer questions real people are actually asking. Publish original data. Build authority over a period long enough for it to be worth something. Contribute what you have personally seen, because first-hand experience is the one input a model cannot produce on its own.

What 400 thin pages did to one brand's AI citations

A large D2C brand came to us after running programmatic SEO to an extreme. Almost 400 pages, all extremely thin, not one carrying original information. Their daily AI citations have been falling ever since, at roughly 60% month on month. That is a severe decline and recovery will take a while. That figure is our own measurement of their citations, not a published study.

The mechanism is resource allocation, not detection. Dan Taylor, writing in Search Engine Journal in July 2026, set out what Google weighs when deciding how much crawling a site earns: perceived inventory against what it judges useful, demand for the topics, and enough domain popularity to justify the processing cost. Publish hundreds of thin pages and you may get burst-crawled, then throttled when the authority does not support the scale.

That explains the pattern people misread. New content gets a temporary freshness advantage, so month one looks like a win. When freshness fades, each page stands on its own, and a page that adds nothing accumulates none of the signals that keep it indexed.

Does Google penalise AI content? No

Ahrefs pulled one million pages from the top ten of 100,000 SERPs collected in June 2026. Fully AI-generated pages do rank at the top: 5.3% of positions one to three came back at 100% AI content. No filter is keeping machine text off page one.

The gradient still runs against volume. Pages under 50% AI content took 82.2% of top-three rankings. Indexation fell from 49.28% for low AI pages to 40.35% for the heaviest. Across two Search Console panels from June 2025 to June 2026, low and moderate AI pages earned two to three times the impressions of high AI pages.

Ahrefs' conclusion is that Google responds to quality, not authorship, and AI use happens to correlate with quality dropping. I agree, and I would put it more plainly. The correlation exists because the people generating at volume are the people putting in no effort.

One caveat worth stating. No study has isolated AI-drafted, human-edited content and measured it. Graphite says in writing it did not test that workflow. The gradient suggests heavy human rewriting moves a page into the better part of the curve, but that is an inference, not a proven result.

The tactics Google says you can ignore

The most useful part of Google's July 2026 guidance is the list of things it says do not work. All of it is scoped to Google Search only.

• llms.txt and similar files. Google states you do not need llms.txt to appear in Google Search including its AI features, because Google Search does not use them, and that maintaining one neither helps nor harms visibility there.

• Chunking content. No requirement to break content into small pieces, and no ideal page length.

• Rewriting for AI systems. Google's systems understand synonyms and meaning without exact keyword matches.

• Chasing mentions. Google says pursuing inauthentic mentions is not as helpful as it might seem.

• Structured data as an AI requirement. Not required for generative AI search, though still worth having for rich results.

One item on that page is a real requirement. To appear in Google's AI features, a page must be indexed, eligible for a snippet, and the site included in Search generative AI features in Search Console.

On llms.txt specifically, we are in agreement with Google. It is a nice to have. We implement it for clients because it does no harm and may help elsewhere, but it does not move the needle and it does not belong near the top of your priority list. Anyone selling it as your route into AI Overviews is selling you something Google has already said it ignores.

GEO and SEO are not separate disciplines

Here is where I disagree with my own industry. People sell GEO and SEO as unrelated fields with two separate packages. Ask a provider to explain each in detail and then articulate the difference. Most cannot.

The foundations are shared and non-negotiable for both: strong technical architecture, content optimised for humans and bots together, and proper JavaScript handling and server-side rendering so the content exists when something arrives to read it. An engine cannot cite a page it cannot render.

The real difference sits in four places:

1. The kind of content you write. Content that answers questions people actually ask, carrying information the model cannot generate itself.

2. The format behind it. Structure decides whether an engine extracts a clean answer from you or gives up and uses someone else.

3. The strategy behind it. You move off keywords and onto queries, which means understanding fan-out, where one question gets split into many sub-queries answered by sources that never ranked for the original phrasing.

4. The surfaces you cover. You identify the company-owned surfaces worth leveraging beyond the website and build unified information across all of them.

That combination is the differentiator. Anything short of it is SEO with new vocabulary and a second invoice.

What we turn down

We decline programmatic SEO more than anything else. We tell every company that we are here to build a consistent machine, and pSEO is the opposite of that.

We also decline companies who will not see a visible return. Sometimes the customer base is too small for AI visibility to move anything. Sometimes those customers are not using AI search to make the decision. We say so and point them at outbound, because that is where their leads actually are.

Ask any provider you are evaluating what they have turned down. One who has never turned work away has not developed a view about who this works for.

How to tell whether an agency is using AI well

Every agency uses AI. The question is where they stop.

1. Ask what they will not automate, and why. A provider who cannot name a boundary has not thought about one.

2. Ask them to explain GEO and SEO separately, then state the difference. If they cannot get past vocabulary to strategy, format and surfaces, you are buying the same service twice.

3. Ask how a statistic gets into a published piece. The answer should describe tracing a figure to the organisation that produced it, not to whoever last repeated it.

4. Ask where their dashboard numbers come from. Google states that third-party tools have no access to its ranking data and cannot guarantee performance. Every tool metric is an estimate.

5. Ask what volume they publish and why that number. Google says a high quantity of pages does not make a site higher quality. A page count with no argument behind it is inventory.

Where you actually stand right now

The prior question to all of this is what AI engines currently say about you, which most companies have never checked.

We run a free AI visibility audit that measures it. Whether ChatGPT, Gemini, Perplexity, Google AI Overviews and Copilot name you when your buyers ask, what they get wrong, who they recommend instead, and which sources they pull from. Read what an AI visibility audit covers, or get in touch and we will run yours.

^{[1][2][3][4][5][6]}