How to use AI for SEO: a practical workflow

Use AI for the parts of SEO that are research, structure and repetition: keyword research with real data, briefs, outlines, first drafts, titles, structured data and reading your Search Console numbers. Keep a person in charge of the facts, the experience and the final say, because Google does not penalize content for being made with AI; it penalizes pages that add nothing, however they were made.

Most guides to AI for SEO are lists of tools. This one is a workflow, in the order I actually use it, for a small business owner or whoever does their marketing. It covers what Google has said about AI content, where AI is reliable and where it quietly is not, the prompts I would start with, and how I use the same process on this site. I link the source for every fact.

Does Google penalize AI content?

No. Google's position has been the same since February 2023: "Appropriate use of AI or automation is not against our guidelines," and "Using AI doesn't give content any special gains. It's just content." (Google Search Central (opens in a new tab)). Its current help page goes further and names the good uses: "Generative AI can be particularly useful when researching a topic, and to add structure to original content" (Google (opens in a new tab)).

The line Google draws is about purpose and scale. Since March 2024, its spam policies have banned scaled content abuse: "when many pages are generated for the primary purpose of manipulating search rankings and not helping users." The first example Google gives is "using generative AI tools or other similar tools to generate many pages without adding value for users" (spam policies (opens in a new tab)). Site owners reported manual actions under the new policies the day after Google announced them (Search Engine Roundtable (opens in a new tab)).

The ranking studies agree once you read them carefully:

• AI use alone does not predict rank. Ahrefs checked about 600,000 top-ranking pages and found 86.5% contained some AI-generated text; the correlation between how much AI a page used and where it ranked was 0.011, effectively zero (Ahrefs (opens in a new tab)).
• Pages made mostly by AI are rare at the very top. Semrush looked at 42,000 blog pages in Google's top ten and found the number one result was far more likely to be human-written (80.5%) than AI-generated (10%), with the rest mixed or unclear (Semrush (opens in a new tab)).

All of these studies rely on AI detectors, which are educated guesses, so treat them as direction, not proof. The direction is consistent: AI-assisted pages with real human input rank fine, and pages churned out by AI with little added rarely win.

What is AI good at in SEO, and what is it bad at?

Here is how I split the work.

The last row is the whole game. Google's quality raters are told to give the lowest rating to pages where almost all of the main content is "auto or AI generated" with "little to no effort, little to no originality, and little to no added value" (Search Quality Rater Guidelines (opens in a new tab)). The same guidelines say AI use "alone does not determine" a page's rating. What you add is what counts.

A practical AI SEO workflow, step by step

1. Connect AI to real data first

A chat assistant on its own has no search data. Ask it for keyword volumes and it will produce numbers that look precise and are made up. Connect it to a real source, or bring the data yourself.

The easiest way now is an MCP server, a connector that lets an assistant like Claude or ChatGPT pull live data from another tool. Ahrefs and Semrush both offer one, and Google offers one for Analytics. Without them, export a report and paste it in.

2. Research keywords and group them into topics

With data connected, AI is excellent at the slow part of keyword research: sorting hundreds of phrases into the handful of topics a page can actually cover, and spotting the questions people ask. A starting prompt:

Here are the keywords my site could target, with monthly searches and difficulty.
Group them by the search intent behind them, not by shared words. For each group,
give the one page that should answer it, the main keyword, and the questions a
searcher would expect it to answer. Flag any group that my existing pages already
cover: [list your pages].

Check its groups against the real results page. If Google shows product pages for a keyword, a blog post will not win it, however well it is written.

3. Read what already ranks, and find the gap

Before writing, look at the pages you have to beat. AI can read the top results and summarize what each covers, what each misses and who it is written for. Ask it to be specific about what is missing, because that gap is your reason to exist:

Here is the text of the top five pages for "[keyword]". For each, list the
questions it answers, the evidence it gives, and what a business owner would
still not know after reading it. Then list what none of them cover.

When I researched this guide, the gap was obvious: none of the top pages quoted Google's actual policy on AI content, and none gave an ordered workflow, only lists of what AI can do.

4. Gather facts from primary sources

Use AI to find sources, not to be one. Have it collect each fact with a link to the original: the vendor's documentation, the study itself, Google's own page. Then open the links. Even with web search switched on, OpenAI's own testing found at least one major factual error in about 5% of its best model's answers; with no web access, it got about 40% of a set of hard factual questions wrong (GPT-5 system card (opens in a new tab)). In an independent study of more than 3,000 answers to news questions from ChatGPT, Copilot, Gemini and Perplexity, journalists found a significant problem in 45% of them (EBU and BBC (opens in a new tab)).

A useful rule: a figure without a link to where it came from does not go in the draft.

5. Outline and draft, then add what only you know

Let AI build the outline from steps 2 to 4: answer first, then the questions in the order a reader needs them. It can write a first draft from that outline. Then do the part it cannot: add your experience, your numbers, your examples and your opinion, and cut every sentence that could appear on anyone's site. Google asks three questions of any page, "Who, How, and Why": who made it, how, and whether it was made to help people or to attract search visits (Google (opens in a new tab)).

Google does not require you to disclose AI use; it suggests doing so "when it would be reasonably expected," and it advises against giving AI the byline (Google Search Central (opens in a new tab)). Put a real person's name on the page, and make sure that person stands behind every word of it.

6. Test the draft against the pages it has to beat

This is the step most people skip, and AI makes it cheap. Give an AI model your draft and the top-ranking pages, with the names removed so it cannot tell which is yours, and ask it to score each one as the person searching would. Ask it to quote the passages behind every score, so you can see what it is reacting to. Then fix what it finds and run it again.

A starting prompt:

You are someone who searched "[keyword]" and needs to make a decision today.
Read documents A, B and C. For each, write the case against it first. Then
score it out of 10 for answering the question, specifics, trustworthiness and
ease of reading, quoting two passages behind each score. Rank them.

Run it more than once. The same model can score the same page several points apart on different runs, so one pass is an opinion, not a verdict.

7. Let AI draft titles, descriptions, alt text and structured data

These are repetitive and rule-bound, which suits AI, as long as you check the output:

• Titles and meta descriptions: Google warns against "repeated or boilerplate text" in titles (Google (opens in a new tab)), which is exactly what bulk AI titles drift into. It also writes most snippets from the page itself, so a meta description is a suggestion, not a guarantee.
• Alt text: describe what the image shows and why it is there, not a list of keywords.
• Structured data: AI writes schema markup well, and also invents fields. Test every piece in Google's Rich Results Test (opens in a new tab), and never mark up anything a visitor cannot see, which Google treats as spam (structured data policies (opens in a new tab)).

Google's rule on AI content covers all of this: it asks for accuracy "especially when automatically generating the content," naming titles, meta descriptions, structured data and alt text (Google (opens in a new tab)).

8. Check every fact again before it goes live

Re-check each figure, quote and product detail against its source, just before publishing. Things change: a report gets a new edition, a feature moves, a link dies. AI can do most of this pass (open each source, confirm the sentence still matches), and you review what it flags.

9. Measure, with AI reading the numbers

Search Console is where AI earns its keep week to week. Export your Performance data (the interface shows up to 1,000 rows; the API gives more) and ask what changed, which pages are gaining impressions but not clicks, and which questions you rank for without a page that answers them. Two tips:

• Filters use RE2 regular expressions (Google (opens in a new tab)), which AI writes well but often with lookaheads, a feature RE2 does not support. Test the filter before trusting the result.
• Watch AI search separately. Search Console now reports impressions in AI Overviews and AI Mode (Google (opens in a new tab)), and Google Analytics has an "AI Assistant" channel for visits from ChatGPT, Gemini, Copilot and others (Google Analytics (opens in a new tab)). Visits from Google's own AI Overviews still count as organic search.

10. Refresh on a schedule

Put anything likely to change, like prices, plan limits and product features, in one clearly dated place on the page, so an update means editing one box rather than rewriting the argument. Then have AI re-check those boxes every few months. Do not just change the date on the page: Google specifically warns against updating dates when "the content has not substantially changed" (Google (opens in a new tab)).

What does it cost, and how much time does it save?

The tools are cheap. The time is the real cost, and it moves rather than disappears.

So a small business can run this whole workflow for about $20 a month using free Google data, or about $150 a month with live keyword and competitor data connected.

As for time, nobody has measured how much AI saves on SEO work specifically. The best controlled experiment on similar work, mid-level professional writing tasks, found people using ChatGPT finished about 40% faster, with better-graded results (Science (opens in a new tab)). Surveys of marketers report bigger savings, but those are people's impressions; in one controlled study of experienced software developers, AI tools made them 19% slower while they believed they had been faster (METR (opens in a new tab)). My experience matches the first study, with a catch: the time AI saves on drafting mostly goes into checking. Budget for that, or the savings come out of your accuracy.

How I use this workflow on my own site

Every guide on this blog goes through the steps above, and I use AI at nearly all of them. Keyword and results data come from Ahrefs through its MCP server. An AI research assistant builds a list of facts, each with its primary source, which I check. I write the draft with AI and edit it until it says what I would say. Then comes step 6, the one I would not skip.

For each post, five AI critics score my draft blind against two real pages from the top of Google for the same search, plus a deliberately hollowed-out copy that should come last. If the hollow copy does not come last, the test itself is broken and I fix the test, not the post. Each critic scores four things out of 100 in total (answering the question, 30; specifics, 30; trustworthiness, 20; ease of reading, 20), quotes the passages behind every score, and argues against every page before scoring it, mine included. The draft has to beat both rivals by at least five points in most passes. My guide to local SEO scored 87 to 90 out of 100 across five critics, against 54 to 64 and 48 to 53 for the two pages it was up against, and the critics' notes still led to four fixes before it went up, including a cost figure two of them said was missing. These are my own tests on my own posts, not an independent benchmark, and an AI critic is still an AI: it tells me what a careful reader might miss, and I decide what to change.

Finally, a day or so before a post goes live, an AI assistant re-checks every claim in it against its source. A post whose facts have not been re-checked does not publish.

None of this replaces reading the post myself. It does mean every guide starts from real data, is checked against the best existing answers, and is re-checked before a reader sees it, which is more than I could do by hand for two posts a week.

What should you never hand to AI?

• Unchecked facts. Every figure needs a source you have opened.
• Your reviews and testimonials. AI-written reviews break Google's policies and, in the US, the FTC's rule on fake reviews, which names "AI-generated fake reviews" among the reviews it bans (FTC (opens in a new tab)).
• Experience you do not have. An AI can describe running a plumbing business. It has not run one, and neither have you if you publish it as your story.
• Publishing at scale. Hundreds of near-identical city or service pages is the textbook case of scaled content abuse, whoever or whatever writes them.
• Your clients' private data, on a consumer plan. See the note below.

Does AI SEO also mean showing up in AI answers?

Increasingly, yes. "AI SEO" gets used for two different things: using AI to do SEO, which is this guide, and getting your business mentioned in the answers ChatGPT, Gemini and Google's AI Overviews give. The second is usually called answer engine optimization, and I cover it in my guide to answer engine optimization. For Google's AI features specifically, Google says there are no special requirements: no extra files, special markup or schema are needed to appear in AI Overviews or AI Mode, and ordinary good SEO is the way in (Google (opens in a new tab)). Be wary of anyone selling you a special file for Google's AI.

For a local business, both kinds rest on the same foundations: a complete Business Profile, genuine reviews and a page for each service, which I cover in my guide to local SEO.

The short version

Use AI for research, structure and repetition, and connect it to real data before you trust any number it gives you. Write the draft with it if you like, then add what only you know and cut what anyone could have written. Test the page against the ones already ranking, check every fact against its source before it goes live, and keep private client data off consumer plans. Google does not care whether AI helped. It cares whether the page is worth finding.