SEO Agency AI: Which Jobs to Hand to AI and Which to Keep Human

An SEO agency can hand AI the first pass on crawl issue grouping, keyword evidence tables, content brief outlines, report narratives and outreach drafts. Keep humans responsible for diagnosis, search intent, factual claims, client strategy, relationship decisions and final publication. Start with one recurring deliverable whose inputs and approval owner are already clear.

For a small agency, the useful unit of delegation is a step, not an entire client account. A model can sort 200 crawl warnings into a review queue; it cannot tell whether a blocked page is deliberately private without understanding the site. That distinction should appear in the brief, the review record and the client-facing report.

Follow me on Instagram@sagnikteaches

Draw the boundary around evidence and approval

Put every recurring job into three columns: source evidence, AI preparation and human decision. Source evidence means the crawl export, analytics extract, live page, client product information or agreed campaign objective. AI preparation means classification, comparison, drafting or finding contradictions in those inputs. The human decision is the action a named specialist approves. This prevents a polished paragraph from becoming a substitute for an inspected URL.

Connect on LinkedInSagnik Bhattacharya

The following split is a starting policy for a five-person team. Change the owner names to match your staff, but leave one accountable reviewer on each deliverable. A reviewer should be able to open the source behind every important claim without asking the person who ran the prompt to reconstruct it.

Subscribe on YouTube@sagnikteaches

Do not let a tool's severity label silently become your agency's priority. Semrush's SEO and SEO + AI Search plans bundle Site Audit and Position Tracking with AI visibility reports, and Ahrefs plans pair Site Audit and Rank Tracker with Brand Radar for tracking AI prompts (Brand Radar is also sold on its own). Their outputs can supply evidence, but the client decision still depends on the affected page, the business goal and the effort to fix it. If the account has no trustworthy source export, pause the AI step and repair the measurement first.

Turn a crawl export into a defensible audit queue

Begin with a crawl date, project domain, issue type, URL, status and any relevant template. Exclude URLs you are not permitted to share with the chosen AI tool. Ask for grouping rather than a final diagnosis. A sensible output has one row per issue family, a sample URL, an estimated affected count, a proposed verification and a confidence label. The team then checks a sample of pages from every group, especially any proposed high-priority item.

For example, an illustrative export says 43 pages have a missing meta description and seven have a noindex directive. AI might call the seven noindex pages urgent. The lead opens all seven: five are internal search results and two are staging copies. None should be indexed. The correct action is to document the intentional exclusion, then inspect whether important pages are missing from the crawl. A severity score alone would have created unnecessary development work.

Another illustrative audit finds 18 product pages returning a redirect after an old catalogue migration. AI groups them as one issue, but the specialist checks whether the destination retains the same product and whether internal navigation still points at the old address. If 16 redirect cleanly and two land on a generic category, fix those two first. This is a narrower, testable client recommendation than “repair all redirects”.

Keep a tiny evidence log: crawl date, tool export name, affected URLs, manual sample, decision, assigned owner and recheck date. The agency-specific AI search audit checklist for agencies is useful when a client also asks whether AI crawlers can access their content. For a general site owner learning the basics, the question is different; this page concerns who in the agency approves the work.

Cluster queries, then inspect the searcher’s job

AI is good at proposing groups from a query list. Give it the exported terms, current ranking URLs, the client's product catalogue and a rule to leave uncertain terms separate. It should produce a proposed page target, queries, likely intent and unanswered question for each group. It should also mark clashes where two existing pages appear to answer the same query. Do not ask it to invent search volume or ranking difficulty.

Suppose a tutoring centre client has the illustrative queries “online maths tutor”, “maths revision lessons” and “maths homework help”. A model may merge all three because the words overlap. The strategist checks the actual results and client offer: a parent seeking regular lessons may need a service page, while a student asking for homework help may need a policy or resource page. Keep the two purposes distinct unless the live evidence supports one page. Record the rejected cluster and the reason so the same mistake does not recur next month.

Before approving a new page, inspect the current site for a suitable destination. A proposed brief is wasteful if the page already exists and needs clearer copy or internal links. The agency keyword clustering method goes deeper into validating groups before sign-off; here the managerial point is to assign the sign-off to a strategist, rather than accepting a model's tidy spreadsheet as a content plan.

Make briefs from verified inputs, not plausible filler

A brief should tell a writer what the reader needs, which claims are supported, what the client can genuinely offer and what remains unknown. AI can assemble those pieces faster if the editor supplies an approved source pack. Give it a product sheet, recorded client answers, page purpose, target questions and a list of claims that require a citation. Make “unknown” an acceptable output. Forbid invented customer examples and prices.

Here is a reusable prompt skeleton for ChatGPT or whichever assistant the agency has approved. Replace the bracketed fields with material the agency can lawfully use. The output is a draft for an editor, not a publication instruction.

Purpose: [one sentence on the searcher's task]
Inputs: [approved product facts, source URLs, existing page]
Create a brief with: reader question, proposed sections,
evidence for each claim, unanswered questions, and pages to link.
Mark any unsupported claim NEEDS CLIENT CONFIRMATION.
Do not invent prices, testimonials, outcomes or search metrics.

An illustrative response for a landscaping company might suggest a section called “How long does a garden redesign take?” and claim completion within two weeks. The client material only states that an initial visit usually takes one hour. The editor deletes the two-week claim, asks the client how scheduling depends on scope and season, and changes the section to explain the consultation and quote process. The model found a useful reader question; it did not supply the missing fact.

Brief quality can be checked with five tests: every section serves the page purpose; each factual claim has a named source; client-specific information has been confirmed; the writer can see what is out of scope; and the brief names the person who approves the finished page. Google Search Central advises producing helpful, reliable, people-first content and has separate guidance for generative AI content. Use those documents as an editorial reference, rather than treating a generated outline as a ranking formula. The official generative AI content guidance is worth including in the editor's reference pack.

Write and edit pages with a source trail

For page copy, a model can suggest alternative headings, remove repetition, turn approved notes into a plain-language draft or list questions the draft fails to answer. The writer still needs to add expertise and inspect every consequential statement. A polished draft can be thin: it may repeat the search query in six ways while saying nothing about the client's actual process. Judge the answer a visitor receives, not the fluency of the sentences.

Consider an illustrative removals firm page. The approved fact sheet says a written survey is required before a fixed quote. AI drafts “Get an instant fixed price for every move”. That is attractive copy but conflicts with the process. The writer changes it to “Send the key details; we will arrange a survey before confirming a fixed quote,” then asks the operations manager whether exceptional cases need a different path. This is a factual correction, not a style preference.

Make a red-line habit for drafts: highlight sentences that make promises, quote figures, mention availability, describe eligibility or tell customers what will happen after they enquire. Check these against source material line by line. Keep a copy of the approved input and the final edited page. If a client later challenges a claim, the agency should know which source supported it and who approved publication.

AI is also useful as a sceptical reader. Ask it, “Which claims in this draft cannot be proved from the attached material?” Give it only the material already cleared for processing. A human checks the flagged claims and looks for omissions the model missed. Never let the model grade its own draft as accurate without external evidence.

Report changes without inventing a cause

Monthly reports are a strong early use case because the agency already has repeatable inputs and an accountable account lead. Export the same metric definitions for the current and comparison periods. AI can calculate differences, draft a short explanation of what changed, identify anomalies and prepare questions for the analyst. It should not say that a new page “caused” a traffic rise simply because the dates line up.

In a small illustrative report, non-brand organic clicks move from 1,200 to 1,380, while enquiries move from 46 to 43. AI writes, “Traffic rose 15%, so the campaign generated more leads.” The first calculation is correct: 180 divided by 1,200 is 15%. The second claim is false. The account lead rewrites it: “Non-brand clicks rose 15%, while recorded enquiries fell by three. We will check whether the increase came from informational pages and whether enquiry tracking changed.” That is useful client communication because it separates observation from hypothesis.

ChatGPT can do the arithmetic if the account lead uploads the reconciled export; OpenAI's pricing page lists data analysis on the Go, Plus and Pro plans. Ask it to show the working for each figure, then recheck two by hand. In one illustrative run it reports enquiries down 6.5%, which is right (a fall of three from 46), but labels the period “last quarter” when the export covers one month. The account lead fixes the label and adds the date range before the commentary reaches the client.

A second illustrative report shows three landing pages with fewer clicks, but one had a tracking tag removed mid-month. An AI summary that only sees the export may describe a performance collapse. The analyst checks the measurement change before discussing rankings. The monthly SEO reporting workflow covers the full report process; the boundary here is that the account lead owns interpretation and promises to the client.

Use a reporting check sheet with the period, source, filter, metric definition, calculated difference, exceptional events, verified explanation, action and owner. If you cannot establish an explanation, say what the team will investigate and when. A client can act on a well-framed uncertainty. They cannot act safely on an invented causal story.

Keep outreach personal for a real reason

AI can shorten a draft email, suggest a clearer subject line or adapt an approved pitch to a recipient whose site has been inspected. It cannot create the relationship or a valid relevance reason. Before drafting, the outreach owner should record the page being contacted, the exact reason the client's resource helps its audience, any prior contact and the offer being made. Approval should include a check that the message matches the recipient's page and the agency's contact policy.

An illustrative yoga studio has published a careful accessibility guide. A generic AI outreach email says the recipient “loves wellness content” and proposes a link exchange. The owner finds that the recipient's page is about venue access, not general wellness, and that the studio guide lacks transport details. The right next action is to improve the guide and identify a genuinely relevant contact, not to send a more personalised version of the same weak pitch.

For a legitimate contact, a draft might say: “Your venue access page explains step-free entry but does not cover how participants can ask about class adaptations. Our published guide has a short question checklist; would that be useful to your readers?” The outreach owner verifies both pages, edits the tone and decides whether to send. The dedicated personalised agency outreach tutorial covers recipient research in more detail.

A six-week pilot for a five-person agency

Use one client with regular reporting and a predictable approval path. The following figures are illustrative assumptions, not a claimed industry benchmark. A five-person SEO agency serves eight clients. For one client's monthly cycle, the team currently spends 4 hours triaging crawl exports, 5 hours assembling a content brief, 3 hours writing report commentary and 2 hours checking it: 14 hours altogether. The agency tests AI on export grouping, brief structure and the first report draft, while the same specialists keep the checks above.

During week one, the account lead records the last two cycles and defines an acceptable output. One audit recommendation must cite a checked URL; every brief claim must trace to the client source pack; every report number must reconcile to the export. In week two, a specialist runs the prompts on historical material without using the outputs for live work. Mark incorrect classifications, fabricated facts and missing caveats. Retain examples of both successes and failures in an internal review sheet.

Weeks three and four use the tool for live first drafts. In this illustration, audit grouping falls from 4 to 2.5 hours including review; brief assembly from 5 to 3.5; report drafting and checking from 5 to 4. The new total is 10 hours, a four-hour difference for that client cycle. The agency then adds 45 minutes for preparing approved inputs and 30 minutes for logging exceptions. Net time released is 2 hours 45 minutes, not four hours. The calculation matters because input preparation and review are part of the work.

In week five, the lead audits a sample of deliverables that passed internal review. Suppose two of ten proposed audit priorities needed a severity change, one of six brief sections carried an unsupported claim, and the report numbers were correct. Those findings call for a tighter audit prompt and stronger brief source references; they do not justify unattended publishing. Week six is a decision: retain the workflow if the team can repeat the quality and time result, revise the weak step, or stop using AI for that step.

Assign one owner to the prompt and version it. When a client changes products, a search tool changes its export columns, or a reporting definition changes, the owner updates the input instructions and tests again. “It worked last month” is not a control. Keep the original source pack, draft, edits and approval in the normal client folder with access limited to the account team.

Set review gates your team will actually use

A review gate should be short enough to fit the agency's normal process. For a technical recommendation, require the checked URL, observed behaviour, impact hypothesis, proposed fix and named approver. For a content brief, require the page purpose, source for each material claim, client confirmation needed and editor approval. For a report, require reconciled figures, a clear difference between fact and hypothesis, and an action the client understands. For outreach, require recipient relevance and a manual send decision.

Use escalation triggers rather than asking reviewers to inspect every trivial word equally. Stop and seek client confirmation when a draft mentions a price, deadline, guarantee, legal claim, customer outcome or product capability absent from the approved pack. Stop when a tool suggests changing indexation, canonical tags or redirects on revenue pages without a specialist's live check. Stop when a report claims causation that the data cannot establish. These triggers catch expensive errors while leaving routine wording changes to the editor.

Access is another gate. A public prompt box may be inappropriate for unpublished client data, credentials or analytics exports. Review the contract, the client's instructions and the tool's current data terms before uploading anything. With ChatGPT, the plan matters: OpenAI's pricing page marks the Free, Go, Plus and Pro plans with “Content is used to train our models” and an opt-out, while its Business page says business data is never used for training. A specialist's personal Plus account is the wrong home for a client's analytics export unless the opt-out is set and the client has agreed. Use the smallest necessary extract, remove identifiers where possible and store prompts with the same discipline as other client working papers. A model does not need access to the whole analytics account to describe a checked table of twelve figures.

Finally, judge the pilot on client value as well as internal speed. Track the proportion of AI suggestions accepted after review, corrections found after delivery, time spent preparing inputs, time spent reviewing, and whether the client can see the evidence behind recommendations. If a faster report creates more clarification calls or a brief produces generic pages, the apparent saving is misleading. Keep the human boundary visible in the deliverable: the agency is accountable for advice, even when AI helped prepare it.

Further reads

• Semrush AI Visibility vs Ahrefs Brand Radar: Which Suits an Agency? — Choose a visibility tool for agency clients.
• AI Schema Markup Generator: Validate Client Structured Data Before Launch — Check structured data before a client launch.
• Scaled Content Abuse: An AI Content Checklist for SEO Agencies — Set a publishing boundary for generated pages.
• Search Console and AI Overviews: What Agencies Can Report to Clients — Explain search reporting limits to clients.
• AI Content Workflow for Agencies: From Brief to Approved Draft — Build a broader agency approval workflow.
• AI SEO Tools for Local SEO: Citations, Reviews and Map Reports — A priced shortlist of local SEO tools for agency clients, compared on citations, review replies and map-pack reports, plus a trial test to run.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Semrush SEO and AI search plans and pricing page; Ahrefs pricing page; OpenAI ChatGPT pricing and ChatGPT Business pages; Google Search Central, Creating helpful, reliable, people-first content and Google Search's guidance on using generative AI content on your website (all checked 4 November 2026).