How to use AI for SEO: what to hand over, what to check, and a weekly workflow.

1. The map: which SEO jobs AI does well, and which it gets wrong.

Most advice on this topic lists twenty things AI can do for SEO and stops. The more useful question is which of those jobs survive contact with a live site. The test is simple: does the job run on text you can hand the model, or on data the model would have to make up? Clustering a keyword export is the first kind. Telling you how many people search for a phrase each month is the second. The same assistant will do both with the same confident tone, which is the whole problem.

Google states the reason in its guidance on generative AI content, last updated on October 1, 2026: generative models don't retrieve facts, but predict a likely sequence of words based on their training data. That is a strength for prose, titles and structure, and a liability for numbers, dates, prices and quotes. The table below sorts the common SEO jobs by that rule. It applies to ChatGPT, Claude, Gemini and Perplexity alike, and to most of the AI SEO tools built on top of them.

Read the table as a division of labor, not a ranking. “Hand over” means the model does the job faster than you and the result is easy to judge by reading it. “Hand over, check” means the job is worth giving it, but one step in the output needs a human or a second pass. “Keep away” means the model has no access to the truth, so its answer is a guess dressed as a fact. The sections below go through each job in the order a real week of SEO automation runs them.

2. Keyword research with AI: ideas yes, numbers no.

Keyword research has a thinking half and a measuring half. The thinking half is where AI earns its keep: give it a product and an audience and it will list the problems buyers search for, the comparison terms, the how-to questions and the jargon, then label each phrase by intent. Give it a messy export of a thousand rows and it will group them into page topics in a minute. For a quick start on the idea side, our blog post idea generator does the same expansion from a single topic.

The measuring half is where it fails. There is no keyword database inside a language model. Ask ChatGPT or Claude how many people search for a phrase and you get a number with a comma in it, stated as calmly as a fact, and nothing behind it. Search-enabled modes do not fix this: a web search reads pages about keywords, not the search engine's own counts. The figure that decides whether a page is worth writing is exactly the one the model cannot know.

The fix is to bring the numbers in and forbid the model from adding any. Export the last 28 days of queries from Google Search Console, plus a keyword tool export if you have one, paste them in, and write the rule into the prompt: use only the numbers in this file, never estimate volume, mark any topic an existing page already covers. Then let it cluster. Our guide to using ChatGPT for SEO has the full clustering prompt, and the Claude for SEO guide shows how to pipe Search Console data in through a connector instead of pasting it.

3. SEO with AI for briefs and drafts.

Briefs are where SEO with AI saves the most time for the least risk. A search-enabled assistant (Perplexity, ChatGPT with search on, Claude with web search, or Gemini) can read the pages ranking for your topic, list the questions they answer, the angles they share and the gaps none of them cover, and hand back an outline with the sources it used. That is an hour of tab-switching done in a few minutes. Save the source list: it becomes the evidence pack the draft and the fact check both work from.

Drafting comes next, and the method matters more than the model. Draft one section at a time from the brief, paste the sources the section relies on, and tell the model to use only figures that appear in them. One giant “write me a 2,000 word article” prompt produces the generic, hedge-everything prose readers skip and an invented statistic in every third paragraph. Section by section, with sources, it produces a draft you can edit. The guide to writing WordPress posts with AI walks through the same loop inside the editor.

Is any of this allowed? Yes. Google's February 2023 post on AI-generated content says appropriate use of AI or automation is not against its guidelines, and that its focus is the quality of content rather than how it was produced. The line it draws is elsewhere: content generated primarily to manipulate rankings breaks its spam policies, and its current guidance warns that generating many pages without adding value for users may violate its policy on scaled content abuse. A useful, checked article written with AI is fine. A hundred unchecked ones are the pattern the policy names. Our AI SEO writer page covers what a draft needs before it counts as useful.

4. Fact-checking: the step that makes AI content publishable.

Every AI draft contains claims nobody checked: a percentage from a survey that does not exist, a quote attributed to the wrong person, a product limit that changed last year. Asking the same chat “are you sure?” does not catch them, because the model answers that question the same way it wrote the draft. What works is a separate pass with one job: list every factual claim in the draft next to the quote from a saved source that supports it, and flag every claim that has none.

Each assistant gives you a tool for this, and none of them is a guarantee. Claude's own documentation on reducing hallucinations recommends asking for a supporting quote for each claim and retracting any claim without one. ChatGPT's help center warns that web search citations can be incomplete, outdated or incorrect, so open them. Gemini has a Double-check response option that runs a Google Search for content likely similar to or different from each statement. Use whichever you have, then read the flagged list yourself and cut or soften every claim the sources do not back.

The check covers more than body copy. Google's guidance applies the same review to title elements, meta descriptions, structured data and image alt text, because those show up in search results too. A meta description promising a feature the product does not have is the same problem as a wrong statistic in paragraph four. This step is the one most teams skip, and it is the one our AI fact checker exists for.

5. Titles, internal links, schema and audits.

Titles and meta descriptions. These are the easiest wins. Ask for ten title variants under 60 characters with the keyword near the front, and five meta descriptions between 120 and 155 characters, then pick. Our free SEO title generator and meta description generator do this with the length limits built in. Read the winners against the page: a title that promises more than the page delivers costs clicks you would otherwise keep.

Internal links. AI is good at matching a new article to related pages and writing natural anchor text, but only from a list of URLs you supply. Without one it invents plausible paths like /blog/refund-guide that return a 404. Paste your sitemap or a crawl export, ask for three to five links per article with the sentence each one goes in, and check every URL resolves. The internal linking tool does the matching against your real pages.

Schema markup. Paste the visible page and ask for Article, BreadcrumbList or Product JSON-LD and you get valid markup in seconds. The two rules are Google's: everything in the markup must be visible on the page, and the result should pass the Rich Results Test before it ships. Our schema markup guide covers which types matter and what markup does not do.

Audits. A chat assistant reads a crawl export, a Search Console coverage report or a page of raw HTML and turns it into a prioritized fix list faster than you can. A coding assistant goes further: Claude Code with SEO skills can check every route in a codebase for missing titles, canonicals or schema and open the fix. What none of them do is crawl a large site or see how Google renders it, so start from a crawler and work through a technical SEO audit checklist with the model as the reader, not the crawler.

6. The other direction: letting AI crawlers read your site.

Using AI for SEO has a second meaning: being found by the assistants your buyers now ask. That depends partly on a file you already own. Each AI company runs separate crawlers for model training and for search, and a robots.txt written in 2023 to “block AI” often blocks the search crawlers too, which can take the site out of those answers. Here is what each one does, from the operators' own documentation.

Two details from that documentation change how to write the file. OpenAI says robots.txt changes for OAI-SearchBot take about 24 hours to apply, and Perplexity says up to 24 hours for its crawlers, so test a change a day later, not a minute later. And Google-Extended is not a crawler at all: Google says it has no separate user agent and does not affect inclusion or ranking in Google Search, so blocking it keeps content out of Gemini training and grounding without touching your search traffic. The ClaudeBot guide has copy-paste robots.txt rules for each case.

One more trap: in a December 2024 study of AI crawler traffic on its network, Vercel and MERJ found that none of the major AI crawlers rendered JavaScript. A page whose text only appears after a script runs was blank to GPTBot, ClaudeBot and PerplexityBot in that test, even with robots.txt letting them in. Getting cited once they can read you is its own discipline, covered in our guides on how to rank in ChatGPT and answer engine optimization, and on the LLM SEO page.

7. SEO using AI: a weekly workflow, step by step.

Put the map together and you get a loop you can run every week with any of the four assistants. The rule running through it: real numbers go in at the start, sources are saved in the middle, and nothing ships until each claim has one. If you keep a standing SEO checklist, this slots in as its content section.

1. Pull real data. Export 28 days of queries from Search Console, plus a keyword tool export if you have one. These are the only numbers the week uses.
2. Cluster, then pick one topic yourself. The model groups the queries and marks what existing pages already cover. You pick, because you know which topic sells.
3. Brief from the live results. A search-enabled assistant reads the top pages and returns an outline, the gaps, and the source list. Save the sources.
4. Draft section by section. Brief plus sources in, one section out, figures only from the sources. Keep your own examples, opinions and product details in. Those are what the model cannot invent.
5. Check every claim. A separate pass lists each claim with its supporting quote. Reword or cut whatever has none.
6. Finish from real inputs. Titles and meta from variants, internal links from your pasted URL list, JSON-LD from the visible page, validated.
7. Publish, then measure. Confirm the AI search crawlers are allowed, publish, and look at the page in Search Console after four weeks before you plan the next round.

Save the instructions instead of retyping them. ChatGPT and Claude both have Projects that keep your brand notes, rules and past briefs in one place, so every chat starts from the same context, and Claude Code can store each step as a skill. If you would rather buy the loop than build it, our roundup of free AI SEO tools covers the single-job options and the ChatGPT SEO tools list covers the add-ons that run inside the chat window.

A worked example: one week, one article.

Harbor Desk is a fictional helpdesk app for small Shopify stores, on the reserved harbordesk.example domain. The export, prompts, logs and hours below are invented to show the method. Its founder runs the workflow above for one week with a chat assistant and Search Console, starting from the query export.

SEARCH CONSOLE · QUERIES · LAST 28 DAYS · HARBOR DESK (FICTIONAL)
query                                  clicks   impr    pos
shared inbox for shopify                    4    880   14.2
shopify helpdesk                            2    610   21.7
canned responses shopify                    0    430   18.9
how to handle refund requests email         0    390   27.4
refund request email template               0    270   31.0
... 207 more rows

Step 2 is where the first catch happens. The clustering prompt forbids estimates, and the model still adds one: a monthly search figure for the refund cluster that exists nowhere in the file. Because the rule is written into the prompt, the extra number is easy to spot and delete. The clusters themselves are good, and they point at a clear gap: 18 queries about refund request emails with no page answering them.

PROMPT  Cluster these 212 queries into page topics. Use only the
        impressions column. Do not estimate search volume. Mark any
        cluster an existing URL already covers.

CLUSTER                          QUERIES   IMPR (28D)   EXISTING PAGE
Shopify helpdesk, shared inbox        31        2,940   /shopify-helpdesk
Refund request emails                 18        1,120   none
Canned responses for Shopify          12          760   none
Order status replies                   9          410   none

RUN 1 added "est. 2,400 searches/mo" to the refund cluster.
The export has no volume column. Number deleted, prompt kept, re-run.

The brief comes from Perplexity reading the top results for “how to handle refund requests by email”, with nine sources saved. The draft is written section by section from those sources plus the founder's own reply templates. Then the fact-check pass lists every claim against the saved sources.

FACT CHECK · "How to handle refund requests by email" · DRAFT 1
claims in the draft                          27
backed by a saved source                     21
reworded to match the source                  3
cut, no source found                          3
  "68% of shoppers never return after a slow refund"      no source
  "Average refund reply time is 4 hours"                   no source
  "Shopify requires refunds within 30 days"                wrong: store policy

Three of 27 claims had nothing behind them, and one was wrong in a way that would have embarrassed a helpdesk company in front of its own buyers. All three read as plausibly as the 21 good ones, which is why the check is a separate pass and not a skim. With titles, three internal links from the pasted sitemap, Article schema and publishing, the week came out like this.

HOURS · ONE ARTICLE · BY HAND WITH AI
export, cluster, pick the topic          1.0
brief from the live results              1.0
draft, section by section                1.5
fact check against saved sources         2.5
titles, links, schema, publish           1.0
total                                    7.0

AI took drafting from most of a day to an hour and a half. It did not make the week short: checking became the longest step, and the whole loop still came to seven hours for one article. At that pace a founder with other work ships about four articles a month. That is the real ceiling of doing SEO with AI by hand, and the honest numbers from our own AI SEO case study show why volume and consistency matter more than any single post.

Common mistakes when using AI for SEO.

1. Trusting the search volume in a chat answer. It is the fastest number to get, which is why it gets used. The model has no keyword database. Take volume from Search Console or a keyword tool, and write “never estimate volume” into every research prompt.
2. Publishing the first draft. The draft reads finished, so it feels finished. It still carries invented statistics, and at volume it is the pattern Google's scaled content abuse policy targets. Run the claim check, then edit in what only you know.
3. Letting the model invent internal links. It writes plausible paths because it has never seen your sitemap. Paste the real URL list every time, and check each link resolves before publishing.
4. Asking a chat window for your rankings. A web search from an assistant is not a rank check for your market, and no assistant can see Search Console unless you paste the data or connect it. Pull positions from Search Console, then ask the model to analyze them.
5. Blocking every AI crawler by default. A blanket rule meant to stop training also blocks OAI-SearchBot, Claude-SearchBot and PerplexityBot, which can take the site out of those answers. Block the training crawlers if you want to, and decide on the search crawlers separately.
6. Retyping the prompt every week. A prompt rewritten from memory drifts, and the output drifts with it. Keep the rules, brand notes and approved examples in a Project or a skill so every run starts from the same instructions.