This month, we were shown a site that had published about three hundred blog posts since last fall, all generated by AI and all "optimized." The traffic graph rose steadily through the winter, then dropped sharply at some point in March.
That drop corresponds to Google's March core update and the new spam policies that accompanied it. One of them has an explicit name: scaled content abuse. Pages produced in bulk to rank, whether by people, by software or by AI, are now clearly targeted.
Meanwhile, other sites are using the same tools, discreetly, to complete their SEO work faster, and they are performing well. The technology is the same; the results are very different. The difference lies in which steps are delegated to the machine.
What does Google actually say about AI-generated content?
This point is frequently misunderstood, so precision is important. In February 2023, Google published guidance stating that it rewards helpful, high-quality content however it is produced, and that using automation, including AI, mainly to manipulate rankings is spam. The guidance is available here: Google Search Central on AI-generated content.
There is therefore no rule against AI. There is a rule against low-quality content. The March policies simply make it clearer that low-quality content produced at scale is treated as spam, regardless of how it was produced.
This leaves considerable room to use AI effectively. Here is the approach we currently recommend for a small business site.
What does a practical SEO workflow with AI look like, step by step?
1. Keyword research: how should AI be used?
Use AI for ideas and dedicated tools for numbers. Describe the business and its customers to an assistant and ask it to brainstorm the questions those customers would actually type. It performs this task well and identifies angles that might otherwise be missed, in English and in French.
What it cannot provide is real search volume, which models readily invent. The ideas should therefore be entered into a proper keyword tool or into Search Console, and the numbers should come from there.
2. Search intent: why verify it yourself?
We recommend keeping this step in-house. For each target phrase, run the search and examine what Google displays: local results and a map, product pages, long guides or videos. This indicates which type of page has a realistic chance. An AI can estimate the intent; the actual results page shows it.
This takes ten minutes per topic, and it prevents writing a 2,000-word guide for a search where people clearly want only a phone number.
3. Outlines: how should an AI outline be used?
Request an outline, then challenge it. In our practice, we usually remove about a third of the sections, because they contain the filler already found in every article on the topic. Then add what only the business knows: prices, local details, how the work is actually done, and the objections customers raise on the phone.
4. Drafting: where do most errors occur?
An AI draft is smooth, even, polite and generic. It contains no opinions, no specifics and no distinctive features. Published as is, it produces a page that resembles a thousand others.
Drafts should be used for structure and for content that is genuinely generic, such as explaining a standard process. Everything else should be written or rewritten by someone who knows the subject. Every fact must be checked, and anything that reads like a brochure should be removed.
We apply a simple test: if a paragraph could appear on a competitor's site without changing a word, it adds no value for your business.
AI-generated text should never be published without a human edit.
5. Meta titles and descriptions: can they be delegated?
Yes, this task can be delegated with confidence. Provide the page, the main keyword and the length limits, and ask for five options. Select one and adjust it. The process is fast and the quality is adequate. The character count should still be verified, as models count poorly.
6. Internal links: who should decide?
The AI identifies the options; you make the choice. Paste a list of your page titles and URLs along with a new article, and ask where the new article should link and which older pages should link to it. The suggestions will be reasonable. You then decide which ones help the reader and which ones amount to link stuffing.
Which SEO steps must remain human?
In summary, these are the steps we do not delegate to a machine:
- Deciding what the business should rank for, based on what it actually sells.
- Reading the actual search results to understand intent.
- Every fact, every price, every date.
- The voice: opinions, local knowledge, and what only you know.
- The final review before publication.
Every other step can be accelerated. These cannot, if the page is to remain effective next year.
Why does this matter more now that AI answers appear above search results?
Last month, Google began displaying AI Overviews above the results in the U.S. This has direct consequences for generic content. If your page says what every other page says, the overview can say it on your behalf, and no one needs to click. The pages that still earn a click, and possibly a citation, are those with genuine substance.
The division of tasks described above is therefore not primarily about avoiding Google's spam team. It is about which kinds of pages retain their value. The machine handles research, structure, variations and repetitive tasks. A person provides judgment, facts and voice.
The greatest risk we currently observe for small local businesses is hiring a provider who uses AI to produce forty blog posts that nobody reads, and presents that as SEO. If your site contains this kind of content and you cannot determine what helps and what hurts, we recommend starting with an SEO audit. In some cases, the most effective step is to remove half of the blog.