Guide · SEO and GEO

How to automate SEO and GEO without publishing spam

AI can write a hundred articles a day, and Google's spam updates keep removing the sites that publish them. This guide shows the other way: two-week sprints where AI and scheduling handle research, drafting, publishing and reporting, and people supply what actually ranks and gets cited in AI answers. Nine stages, example prompts, and a guardrail for each.
Read the 9 stagesSee the sprint plan

The split

What AI should do, and what people must

The rule of thumb: automate the work that moves information around, keep people on the work that creates it or vouches for it. The two highlighted rows are the ones that separate useful content from spam.
JobAI and automationPeople
Topic researchPull and cluster questions from Search Console, sales calls, support tickets and AI answersPick which questions are worth answering
BriefsOutline what ranks and what AI answers already sayAdd what only you know: data, cases, opinion
Expert input (recommended)Transcribe and tidy a 20-minute interviewBe the expert, or get one on a call
First draftDraft from the brief and the transcript, flagging every claimRewrite the parts that sound generic
Fact-check (recommended)List every claim, number and link to checkCheck them and sign the piece with a real name
PublishingTitles, meta descriptions, alt text, internal links and schema as drafts; schedule the postApprove what goes live
DistributionDraft the LinkedIn post, newsletter section and pitchPost as yourself and answer replies
MeasurementWeekly Search Console report, monthly AI answer check, a refresh queueDecide what to refresh, merge or retire
Google's position: how content is made matters less than whether it's original, helpful and made for people.

The sprint

Ten working days, four to six pieces

Each sprint runs the nine stages below. The two human-heavy days are the interviews and the edit. Everything else is mostly automated or AI-assisted.

Two weeks per sprint · measurement runs weekly in the background

  1. D1: Collect questions

    Automated pull and clustering

    Stage 1

  2. D2: Pick and brief

    Score, cap, write briefs

    Stages 2–3

  3. D3: Interviews

    20 minutes per piece

    Stage 4

  4. D4–6: AI drafts

    From brief and transcript

    Stage 5

  5. D7–8: Edit and fact-check

    Named author signs off

    Stage 6

  6. D9–10: Schedule and spread

    Queue posts, draft repurposing

    Stages 7–8

The 9 stages

How each stage works, and what keeps it honest

Every stage says who owns it, how much of it to automate, and a guardrail that keeps the output on the right side of Google's spam policies.

Days 1–2

Plan the sprint

Most AI content fails before a word is written: it answers questions nobody asked, or says what every other page says. Planning is where you stop that.

Mostly automatedMarketing lead

Stage 1: Collect the questions real buyers ask

Why it matters

Keyword tools tell you what is searched, not why. Your best topics are already sitting in your Search Console queries, sales call notes, support tickets and inbox, and in the questions people ask ChatGPT and Perplexity about your category. Pages built on real questions are the ones people, and AI answers, have a reason to use.

What to do

  • Export the last 90 days of Search Console queries, plus anything from your CRM, call transcripts and support desk you can get as text.
  • Ask ChatGPT, Perplexity and Google your buyers' ten most common questions. Note which sites get cited: they're who you're up against.
  • Have AI merge duplicates, group questions by intent, and tag each group with where it came from.
  • Run this on a schedule (weekly or at the start of each sprint) so the backlog fills itself.

Spam guardrail

Every topic in the backlog traces back to a real person's question. If you can't point to who asked it, it doesn't go in.

Prompt
Here are questions from our Search Console queries, sales calls and
support tickets. Group them by the underlying question a buyer is
trying to answer. For each group give:
- the question in the buyer's own words
- how many raw questions it covers and where they came from
- whether it's research, comparison, pricing or how-to
Do not invent questions that aren't in the data.
Human decisionMarketing lead + subject expert

Stage 2: Pick a sprint backlog you can review properly

Why it matters

Google's scaled content abuse policy targets pages produced in volume mainly to rank, however they were written. The August 2026 spam update hit sites with thousands of programmatic and AI pages. The limit on how much you publish shouldn't be how fast AI can write, but how much your experts can review well.

Source: Google: spam policies, scaled content abuse

What to do

  • Score each question on three things: does answering it help sell what you sell, can you say something the top results don't, and how much work is it.
  • Cap the sprint at what your reviewers can check properly. For a small team that's usually four to six pieces every two weeks.
  • Mix new pieces with refreshes of pages that are slipping (stage 9). A refresh is often the cheapest win in the sprint.
  • Give every piece an owner who knows the subject, not just an editor.

Spam guardrail

Ask of every piece: would we publish this if search engines didn't exist? If the only reason is a keyword, drop it.

AI drafts, human finishesPiece owner

Stage 3: Write a brief that says what only you can add

Why it matters

If you give AI a keyword, you get the average of what's already ranking. Google's helpful content guidance asks whether a page offers original information, reporting, research or analysis. A brief is where you decide what that will be before anyone drafts.

Source: Google: creating helpful, people-first content

What to do

  • Have AI summarise the top five results and what ChatGPT and Perplexity currently answer: what they cover, what they skip, what they get wrong.
  • Write the answer in one or two sentences. That goes at the top of the piece, so readers and AI systems get it without scrolling.
  • Add the part nobody else has: your numbers, a client case, a mistake you've seen, a view you'd defend.
  • List the sources you'll cite, and the expert who'll be interviewed.

Spam guardrail

If the "only we can add" section of the brief is empty, the piece isn't ready to draft.

Days 3–8

Produce pieces worth citing

AI does the drafting. People supply what a model can't: experience, numbers from your own business, and a name that stands behind the claims.

AI transcribes and tidiesSubject expert

Stage 4: Get first-hand input on record

Why it matters

This is the step that separates useful content from spam, and it takes 20 minutes. Experience is the hardest thing for a model to fake and the easiest thing for a reader to spot. In the GEO study, adding quotes and statistics to a page raised its visibility in AI answers by up to 40%. Your own expert and your own numbers are the best version of both.

Source: GEO: Generative Engine Optimization (KDD 2024)

What to do

  • Record a 20-minute call with your expert using the brief's questions. A founder, an engineer or an account manager all count.
  • Ask for specifics: a number, a recent example, what they'd tell a friend, what most advice gets wrong.
  • Pull your own data where you can: averages from your projects, survey results, before and after figures.
  • Have AI turn the transcript into clean notes and quotable lines, keeping the expert's words.

Spam guardrail

Every piece has at least one thing from the interview or your data that a competitor couldn't have written.

AI draftsPiece owner

Stage 5: Draft with AI, limited to the brief and the sources

Why it matters

Google says using AI to create content isn't against its guidelines; using it mainly to manipulate rankings is. When the model writes only from your brief, transcript and sources, it saves hours of typing without filling the page with padding and invented facts.

Source: Google: guidance about AI-generated content

What to do

  • Give the model the brief, the transcript notes and the source list, and tell it not to add facts from anywhere else.
  • Ask for the answer first, short sections under question-style headings, and tables where a comparison helps.
  • Have it mark every number and claim with [CHECK] so the editor can find them.
  • Keep your style guide in the prompt so drafts sound like your company, not like a chatbot.

Spam guardrail

The draft has no statistic, quote or claim that isn't in the sources you gave it.

Prompt
Write a draft using ONLY the brief, interview notes and sources below.
- Open with the direct answer in 1-2 sentences.
- Use the expert's quotes word for word, attributed by name.
- Put [CHECK] after every number, date and factual claim.
- If something in the brief isn't covered by the sources, write
  [GAP: ...] instead of filling it in.
- Follow the attached style guide. No filler intros or summaries.
AI pre-checks, human signsEditor + subject expert

Stage 6: Edit, fact-check and put a real name on it

Why it matters

Since January 2025, Google's quality raters are told to give the lowest rating to pages where most of the content is AI-generated with little effort, originality or added value. Editing is where the effort becomes visible: a named author, checked facts, and a voice that sounds like a person who has done the work.

Source: Search Engine Land: raters told to flag low-effort AI content

What to do

  • Have AI list every [CHECK] item with the source it came from. A person confirms each one, and fixes or removes what doesn't hold up.
  • Cut generic sentences. If a line could appear on any competitor's site, rewrite it with a specific or delete it.
  • Check what AI generated around the article too: title, meta description, image alt text and structured data.
  • Publish under a named author with a short bio that says why they know the subject.

Spam guardrail

The named author has read the final version and would defend every claim on a sales call.

Before you publish anything

Can ChatGPT, Claude and Perplexity actually read your pages?

Great content is invisible to AI answers if your firewall blocks assistants or your text only appears after JavaScript runs. The free checker tests it in about ten seconds.

Check my website

Days 9–10, then weekly

Ship, spread and keep it fresh

Publishing is the start. Mentions on other sites and regular refreshes are what keep a page ranking and cited after the first month.

AutomatedMarketing ops

Stage 7: Schedule publishing at a steady pace

Why it matters

Scheduling is safe to automate completely once a piece is approved. A steady rhythm also makes your output easier to review and measure than publishing in bursts, and keeps any one week from looking like a content dump.

What to do

  • Queue approved pieces in your CMS on a fixed cadence, for example Tuesday and Thursday mornings.
  • Let automation add the routine parts: internal links to and from related pages, Article schema, sitemap entry, social preview image.
  • Notify Bing and Copilot through IndexNow, and request indexing in Search Console for the pages that matter most.
  • Check the page renders in server HTML so ChatGPT, Claude and Perplexity can read it (see the agent-ready guide).

Spam guardrail

Nothing is published that hasn't passed stage 6. Automation schedules approved work; it never approves it.

AI drafts, people postFounder, experts, PR

Stage 8: Earn mentions beyond your own site

Why it matters

In Ahrefs' study of 75,000 brands, the number of times a brand was mentioned around the web was the strongest predictor of appearing in Google's AI Overviews, about three times stronger than backlinks. AI answers repeat what many independent sources say about you, so your own blog alone won't get you there.

Source: Ahrefs: AI Overview brand visibility, 75k brands

What to do

  • Have AI cut each piece into a LinkedIn post, a newsletter section and a short pitch for journalists or podcasts.
  • Pitch the original data from stage 4. A number nobody else has is what gets picked up and cited.
  • Answer questions in communities your buyers use, as yourself, linking only when it genuinely helps.
  • Keep your profiles on review sites, directories and partner pages accurate. AI answers draw on them too.

Spam guardrail

People post as themselves. No AI-written community comments, fake accounts, bought reviews or paid posts on other sites dressed up as editorial.

Mostly automatedMarketing ops

Stage 9: Measure every week, refresh what slips

Why it matters

AI assistants lean towards recent pages: Ahrefs found the URLs they cite are on average 25.7% fresher than those in Google's organic results, with ChatGPT the strongest. Meanwhile clicks from Google are falling where AI Overviews appear. Tracking both tells you which pages to refresh and whether you are being cited even when you aren't clicked.

Source: Ahrefs: AI assistants prefer fresher content

What to do

  • Every Monday, pull clicks, impressions and position per page from the Search Console API into a sheet or dashboard.
  • Every month, ask ChatGPT, Perplexity and Google a fixed set of 20 buyer questions and log whether you're cited, and who is.
  • Flag pages that lost more than a fifth of their clicks, or whose facts and prices are out of date, and add them to the next sprint.
  • Refresh properly: new data, new examples, sections that answer new questions. Then update the date.

Spam guardrail

Only change a page's date when its content has really changed. Fake freshness is easy to spot and erodes trust in the whole site.

Weekly schedule
Mon 07:00  Search Console API -> sheet: clicks, impressions, position
Mon 07:15  Flag pages down >20% over 28 days -> refresh queue
1st of mo  Run 20 buyer questions in ChatGPT, Perplexity, Google
           -> log cited (yes/no) and competing sources
Sprint day 1  Refresh queue + new questions -> backlog

On a schedule

What runs by itself

Set these up once and the sprint starts with a full backlog and a refresh list. Any automation platform (n8n, Zapier, Make) or a few small scripts will do.
WhenWhat runsWith
Every MondaySearch Console report and refresh flagsSearch Console API, a sheet or dashboard
Sprint day 1Question pull and clusteringCRM, call transcripts, support desk, an LLM
Tue and Thu (recommended)Publishing approved piecesYour CMS scheduler
On publishInternal links, schema, sitemap, IndexNowCMS plugins or a small script
MonthlyAI answer check on 20 buyer questionsChatGPT, Perplexity, Google, or a tracking tool
QuarterlyPrune or merge pages with no traffic and no purposeSearch Console export, a person deciding
Publishing is scheduled, never automatic: only pieces that passed the pre-publish gate go in the queue.

What to skip

Shortcuts that look like growth and end as penalties

Each of these works for a few weeks on some sites. Each is also exactly what Google's spam policies and quality raters are set up to catch.

  • A page for every keyword, city or product combination

    Templated pages that differ only by a swapped word are what the scaled content abuse policy describes, and what the August 2026 spam update removed from search in bulk. Build a location or comparison page only when you have something different to say on it.

    Source: GSQI: August 2026 spam update case studies

  • Publishing as fast as AI can write

    Hundreds of posts a month with no expert review isn't a strategy, it's a risk to the whole domain. Set the pace by review capacity (stage 2), not generation speed.

    Source: Google: spam policies, scaled content abuse

  • Stuffing keywords for AI answers

    In the GEO study, keyword stuffing did essentially nothing for visibility in AI answers. Quotes, statistics and cited sources did. Write for the question, not the phrase.

    Source: GEO: Generative Engine Optimization (KDD 2024)

  • Fake freshness and fake people

    Bumping dates without changing content, invented author personas, AI comments in communities and bought reviews all undermine trust once spotted. Real names and real updates are what make freshness and mentions count.

    Source: Search Engine Land: raters told to flag low-effort AI content

The pre-publish gate

Eight checks every piece passes before it's scheduled

This is the line between automation and spam. AI can help prepare every item on the list, but a person ticks it. Put it in your CMS as a required checklist, so nothing reaches the publishing queue without it.

The last question matters most. It's close to the test in Google's own guidance on people-first content, and it's the fastest way to spot a piece that exists only for a keyword.

Making sure AI can read what you publish is a separate job. Our agent-ready website guide covers it.

Pre-publish gate

  • Answers a question a real buyer asked (done)Traceable to Search Console, a call, a ticket or an AI answer.
  • Says something the top results don't (done)Your data, a case, an expert view, or a clearer answer.
  • The direct answer is in the first two sentences (done)
  • Every number, quote and claim is checked and linked (done)No [CHECK] or [GAP] markers left.
  • A named author with real expertise has signed it off (done)
  • Title, description, alt text and schema reviewed by a person (done)
  • Readable without JavaScript, and linked from related pages (done)View Page Source shows the text, and at least two internal links point to it.
  • Would we publish it if search engines didn't exist? (needs attention)If the honest answer is no, don't publish it.

FAQ

Questions about automating SEO and GEO

Does Google penalise AI-generated content?

Not for being AI-generated. Google says appropriate use of AI or automation isn't against its guidelines. What it acts on is content produced mainly to manipulate rankings, especially at scale, with little originality or value. The same rules apply to content written by people.

What's the difference between SEO and GEO?

SEO aims to rank in search results. GEO (generative engine optimisation) aims to be used and cited in AI answers from ChatGPT, Perplexity, Gemini and Google's AI Overviews. In practice they reward much the same thing: clear answers, original information, a site AI can read, and a brand other sites talk about. One process covers both.

How many articles should we publish per sprint?

As many as your experts can review properly. For a small team that's usually four to six pieces every two weeks, including refreshes. Two strong pieces beat ten generic ones, both for rankings and for being cited.

Can we fully automate content from start to finish?

You can automate research, drafting, scheduling, internal linking and reporting. You shouldn't automate the decision to publish, the expert input, or the fact-check. Those are what make content worth ranking, and what Google's raters look for.

How long until we see results?

Expect early movement on refreshed pages within weeks and on new pages within two to four months, depending on your site's authority and competition. Mentions on other sites build more slowly. Judge the process after three sprints of data, not one.

How do we know if AI assistants are citing us?

Keep a fixed list of 20 buyer questions and ask them in ChatGPT, Perplexity and Google every month, logging whether you're cited and who is. Several tools automate this. Also look for AI referrers such as chatgpt.com and perplexity.ai in your analytics.

Which tools do we need?

Less than vendors suggest: Search Console, your CMS scheduler, a transcription tool, an LLM such as Claude or ChatGPT, and a sheet or dashboard. Automation platforms like n8n or Zapier connect them. The process matters more than the stack.

Last reviewed . Sources are linked under each stage. Also see Google: optimising for AI features in Search.

Next step

Want the system built rather than read about?

We'll set up the automations around your CMS, Search Console and team: the question pipeline, drafting prompts in your voice, the pre-publish gate, scheduling and reporting. Your experts stay in the loop; the busywork doesn't.
See our automation workBook a content system review