AI for SEO Work: Assist the Analyst, Don't Replace Them

7 October 2026 · The Cresia team

Yes, AI can help with SEO work, as long as it never gets the job of deciding what's true. Search Engine Journal's write-up of one technical SEO's workflow shows a shape that holds up: scripts check the facts, a language model explains them, and a person decides. For teams working on AEO and GEO, that's a sturdier template than doing everything by hand or handing the whole audit to a chatbot.

Key takeaways

Keep three jobs separate: scripts establish facts, a language model explains them, a person decides what to do.
Never let a model be the source of a fact such as a status code, a canonical tag, or whether your brand appears in an answer.
Answer engines favour passages that are easy to retrieve, easy to read and consistent with other sources. Audit for that, not for tricks.
Start with one recurring audit. Write its deterministic checks first and add the model's explanations last.
Measure the workflow by review time and error rate, and measure AI visibility with repeatable prompt sets rather than one-off screenshots.
Avoid autopilot publishing, unreviewed model summaries, and dashboards nobody has agreed to act on.

What the Search Engine Journal piece describes

Search Engine Journal ran a post by a technical SEO who divides audit work into three jobs. Deterministic checks establish facts: is the page indexable, what does the canonical say, what status code came back. A local language model then turns those findings into plain-English explanations. A human reads both and judges what to do about it.

Two details are worth underlining. The checks come first and don't involve a model, so a status code is never a guess. And the model runs locally, which has a practical side effect: audit data stays on your own machine and you aren't paying per token for every explanation.

Neither point is exotic. That's the appeal. Most of what goes wrong when teams bring a chatbot into SEO is that the chatbot is asked to observe, explain and decide all in one step, and nobody can tell which of the three it got wrong.

Why the split matters for AEO and GEO

Classic technical SEO has crisp facts. A redirect chain either exists or it doesn't. AEO and GEO work is blurrier, and that's exactly why the discipline matters more there.

Take a typical question from a head of SEO: why did our product comparison page stop showing up in AI answers? The tempting move is to paste the page into a model and ask. You'll get a fluent, confident paragraph. It may mention thin content or weak authority. None of it is an observation. It's a plausible story.

Now run the same question through the three-job split. The scripts check what can be checked: does the page return 200, is it blocked to any crawler in robots.txt, did its structured data change, is the main text present in the raw HTML or only after JavaScript runs, did the page's last-modified signal move. The model takes that list and writes a plain explanation for the content lead. The strategist looks at it and decides whether to fix rendering, rewrite the opening passage, or accept that the page lost to a better one.

The output is less dramatic and much more useful. You can argue with a checklist. You can't argue with a vibe.

How answer engines choose and cite sources

The details differ by engine and they change often, so treat anything below as a working model and check each vendor's own documentation for crawler names, controls and citation behaviour. The broad pattern is stable enough to plan around.

Most answer engines do some version of the same thing. They interpret a question, retrieve candidate pages or passages from a search index or their own crawl, assemble an answer with a language model, and show some sources alongside it. Several things in that chain are decided before any model writes a word:

Access. If a crawler can't fetch the page, or the content only appears after client-side rendering it doesn't run, the page can't be a candidate.
Retrievability. Passages that state a clear claim near a clear heading are easier to match to a question than a long page that circles its point.
Consistency. A claim that matches what other reputable pages say is safer to repeat than one that stands alone.
Freshness and specificity. For questions about prices, versions or recent events, a page with clear dates and named specifics has an edge over a general one.

Notice how many of those are checkable by a script. Access, rendering, headings, dates and schema are facts. Whether a passage is persuasive, or whether your claim is one a cautious engine would repeat, is judgment. That boundary is the same boundary the Search Engine Journal workflow draws.

What to split into scripts, model and human

Here's one way to divide a recurring AEO audit. The owners are examples; adjust them to your org chart.

Audit question	Deterministic check	Model's job	Human's job	Suggested owner
Can AI crawlers reach the page?	Fetch with relevant user agents, read robots.txt and response codes	Explain blocked paths in plain language	Decide which bots to allow, per policy	SEO lead with legal
Is the answer in the HTML?	Compare raw HTML with the rendered page	Summarise what's missing	Decide whether to change rendering	Web engineering
Is the opening passage quotable?	Extract first 100 words under each heading, flag headings with no direct answer	Draft a plain description of the gap	Rewrite the passage, keep the facts right	Content editor
Is structured data intact?	Validate markup against what's on the page	Explain mismatches	Decide which source of truth wins	Analytics or dev
Do we appear in tracked prompts?	Run a fixed prompt set on a schedule, log the answers and cited domains	Cluster the answers into themes	Judge which losses matter commercially	Growth lead
Are the claims still true?	Diff numbers, prices and dates against the source of record	List what changed	Approve edits and sign off	Subject-matter owner

The middle columns are where most teams under-invest. The scripts are cheap to write once and run forever. The explanation column is where a model earns its keep, because it turns a 300-line diff into three sentences a busy editor will actually read.

Setting it up: a first pass in a week

Don't build a platform. Pick one audit you already run every month and rebuild it in this order.

List the facts you currently eyeball. Status codes, canonicals, headings, schema types, dates, whether a named entity appears. Anything a person copies from a tool into a spreadsheet belongs here.
Script each fact. Output plain rows: URL, check, result, timestamp. No prose. If a check can't be made deterministic, it's a judgment and goes in the human column.
Write the explanation prompt last. Feed the model only the rows, tell it to describe what the rows say and nothing else, and have it label anything it's inferring rather than reading.
Make the human step explicit. A reviewer opens the rows and the explanation side by side, accepts or edits, and records the decision. If the explanation and the rows disagree, the rows win.
Keep a log of corrections. Every time a reviewer fixes the model, note why. After a month you'll know which explanations to trust and which to drop.

A local model suits this because the input is your own crawl data and the task is narrow. You don't need a frontier model to describe a table. You do need the discipline of keeping the task that small.

How to measure it

There are two things to measure, and teams often blur them. One is whether the workflow is working. The other is whether your AI visibility is improving.

For the workflow, track reviewer time per audit, the number of explanations edited or rejected, and the number of errors found after sign-off. If review time falls and post-sign-off errors don't rise, the split is paying for itself. If edits are heavy, the prompt is too ambitious or the rows are too thin.

For visibility, use a fixed set of prompts, run on a schedule, with the answers and cited sources logged. Change the prompts rarely, or you can't compare month to month. Ahrefs has written about the mess in AI search return on investment and about getting unstuck when visibility data piles up; both are worth reading for how they frame the problem, and neither makes the numbers cleaner than they are. Expect variation between runs of the same prompt. Read trends over weeks, not a single bad Tuesday.

If you want a managed view of AI-search and media performance in one place, look at MediaPilot. And when your answers depend on events fired from your own site, get the definitions written down first. A tracking specification is the plain-language record of what each event means, and it keeps the facts column honest.

What not to do

A few habits will quietly undo the benefit.

Don't let the model grade its own homework. If the same model writes a passage and then scores whether it's citable, you've built a mirror.
Don't publish unreviewed summaries. An explanation meant for your editor is not copy for your readers.
Don't chase a single citation. One screenshot of your brand in an answer is an anecdote. One screenshot of your competitor in an answer is also an anecdote.
Don't write for the machine at the expense of the reader. Clear, direct passages help both. Stuffed or oddly phrased text helps neither and is the first thing a human reviewer will cut.
Don't automate the decision. The moment a model is allowed to change a robots.txt rule, a canonical or a price on its own, you have a different, riskier system.
Don't skip the legal and brand conversation. Whether to allow a given AI crawler is a policy choice, not a technical default. Search Engine Land has covered publishers pressing for AI crawlers to be more transparent, which shows the question is live and unsettled.

Where this leaves the AEO and GEO team

The useful stance is modest. Language models are good at turning structured findings into readable explanations and at drafting first versions of things a person will edit. They are poor witnesses. They can't tell you what a crawler saw, and they'll fill gaps with plausible text.

So spend your effort where the leverage is. Make facts cheap to collect and impossible to dispute. Make explanations short enough to read. Keep judgment with someone who knows the business and can be held accountable for the call. If you're setting up the broader operating model around that, the platform overview shows how Cresia frames the pieces.

This will keep changing. Engines add and rename crawlers, adjust how they cite, and test new formats; OpenAI testing visual ads in ChatGPT image generation, as Search Engine Journal reported, is a reminder that the surfaces themselves are moving. A workflow built on checks, explanation and judgment survives those shifts better than one built on a particular tactic, because only the checks need updating.

Frequently asked questions

Can AI replace an SEO analyst for AEO and GEO audits?

Not for the parts that carry accountability. A model can summarise findings and draft options, but it can't observe what a crawler fetched or decide which trade-off suits your business. The workable setup is one where the analyst's time moves from collecting and describing to reviewing and deciding.

What should be deterministic in an AI search audit?

Anything that has a single correct answer you can check: response codes, robots rules, rendered versus raw content, schema validity, dates, and whether a named brand or source appears in a logged answer. If two people could disagree about the result, it's a judgment and belongs with a human.

Do I need a local model, or will a hosted one do?

Either can work. A local model keeps crawl data on your own infrastructure and avoids per-request costs, which suits narrow explanation tasks. A hosted model may be fine if your data policy allows it. Check your own security and vendor terms before sending client data anywhere.

How often should we re-run AI visibility checks?

On a schedule you can keep, usually weekly or monthly for a fixed prompt set, with ad hoc runs after major content or site changes. Answers vary between runs, so look at direction over several cycles instead of reacting to a single result.

Where should a team start if it has no workflow today?

Pick one monthly audit, list the facts a person currently checks by eye, and script them. Add the model's explanation only after the rows are reliable, and put a named reviewer in the loop from day one. If you want help shaping it, you can request a demo.

Sources

https://www.searchenginejournal.com/using-ai-to-assist-with-seo-work-not-replace-the-worker/591045/
https://ahrefs.com/blog/ai-visibility-workflow/
https://ahrefs.com/blog/ai-search-roi/