Short answer: An AI SEO agency spends most of its week on three things — deciding what to target, writing pages that answer those questions better than the current top results, and checking the technical plumbing that lets search engines understand them. Agents handle the volume: clustering keywords, drafting from research, validating schema. A human decides what to skip, edits every published word, and signs the report. If nobody on your account can name the keyword each page is claiming, that is not AI SEO. That is invoicing.
Adaeze runs a three-person conveyancing practice. She had a website built in 2022 by a cousin who was, and remains, very good at graphic design. It is a handsome site. It has never once appeared for the phrase “property lawyer” followed by the name of her city.
In 2024 she hired an agency. They sent a monthly PDF with a chart that went up. When she asked which page had improved, they sent a longer PDF. After eleven months she cancelled, and when she asked what she was left with, the answer was the eleven PDFs.
The thing that annoys her, two years later, is not the money. It is that she still cannot say what those eleven months were spent doing.
Illustrative composite — not a client, not a testimonial
That is the question this piece answers. Not what SEO is — there are ten thousand posts about that, and most of them were written by people being paid by the word to avoid answering. This is what the week actually contains, hour by hour, including the parts nobody puts in a case study because they photograph badly.
The honest version of the job
Search work is roughly thirty separate tasks wearing one trench coat. About four of them are interesting. The other twenty-six are the reason it works.
The interesting four are the ones agencies lead with: strategy, positioning, content angles, competitive analysis. Genuinely important, genuinely enjoyable, and collectively about a day a month. The remaining twenty-six are things like checking that your canonical tags do not contradict each other, noticing that a plugin update quietly added a second H1 to every service page, and confirming that the sitemap still lists routes that exist.
Nobody has ever won an award for fixing a sitemap. It is also the single most common reason a site with good content ranks for nothing.
Monday: the keyword massacre
Every engagement starts with a keyword export, and every keyword export is mostly garbage. This is not a criticism of the tools. It is what happens when you ask a machine for every phrase containing a word.
Pull the data for an accounting firm and you get, in no particular order: people searching for accounting degree salaries, people trying to log in to accounting software, people looking up a government tax portal, three competitor brand names, somebody researching a video game with “tycoon” in the title, and — buried in there — about forty phrases belonging to people who want to hire an accountant this month.
Roughly three rows in four never earn a page. The job is to throw them out loudly, in a document the client can read, with the reason attached. Not because clients enjoy reading skip lists, but because the alternative is an agency quietly targeting “accounting degree salary” for six months and reporting the traffic as a win.
What survives gets sorted by what the searcher actually wants:
- Hire someone now. The phrases that convert. Expensive, competitive, worth every hour.
- What does this cost. The anxiety searches. Almost nobody answers them properly, which makes them the easiest genuine win available.
- I am in trouble. Behind on filings, missed a deadline, got a letter. High intent, low competition, and the pages nobody writes because they are uncomfortable.
- Explain this to me. Low commercial value individually, but they build the topical weight that makes the first three rank at all.
One primary keyword gets assigned per page and logged. It never gets used twice. The most common self-inflicted wound in SEO is three pages on one site quietly competing against each other for the same phrase, and it happens because nobody kept a list.
Midweek: writing against what already ranks
Here is the part people find deflating: before writing anything, you read the competition. All of it. The current top results for the target phrase, start to finish, with a notebook.
You are not reading for prose. You are reading for shape:
- What format is winning — a list, a decision framework, a walkthrough?
- How long are they, roughly?
- What does every single one of them cover?
- What does none of them cover?
That last question is the whole job. In nearly every commercial search, the top results share the same blind spot, and it is almost always one of three things: the actual price, the situation where you should not buy this yet, or the real arithmetic behind a claim everyone else hand-waves.
They skip those because they are uncomfortable to write. Which is exactly why covering them works.
Structural patterns get borrowed. Sentences never do. Reading what ranks teaches you the shape of a good answer; copying it teaches Google that you are a duplicate.
Then it gets written — a direct answer at the top for anyone who wants the conclusion and not the essay, the sections that match what winning pages cover, the section nobody else has, and a genuine FAQ drawn from the questions people actually ask rather than six rephrasings of the title.
Friday: the unglamorous checklist
A page is not finished when the words are done. It is finished when all of this is true:
- The title fits inside what search engines actually render, with the keyword near the front
- The meta description says something specific rather than restating the title
- There is exactly one canonical tag and it points at the right URL
- Structured data is present, valid, and matches what is visibly on the page
- Internal links point at routes that return 200 — no aspirational links to pages that do not exist yet
- Images have alt text that describes something, not “image1”
- The page is in the sitemap
- The keyword is logged as claimed
- It loads acceptably on a mid-range Android phone on a patchy connection
That last one gets skipped constantly by agencies working from Western templates, and it matters enormously in markets where most traffic arrives on a mid-tier handset. A page that takes nine seconds to become useful has, functionally, not been published.
What the machine does, what the human does
“AI SEO” is a phrase doing a lot of work in a lot of marketing, so here is the actual division of labour:
| Agents handle | A person handles |
|---|---|
| Clustering thousands of keyword rows by intent | Deciding which clusters are wrong for this business |
| Pulling and summarising what currently ranks | Spotting the gap worth writing into |
| Drafting against an approved outline | Editing every sentence that publishes |
| Validating schema and metadata on every route | Deciding what the schema should claim |
| Watching positions daily and flagging movement | Explaining what the movement means and what to do |
| Assembling the monthly report | Signing it, including the parts that went backwards |
The pattern is consistent: machines are excellent at volume and consistency, and reliably confident when they are wrong. Every place where being confidently wrong is expensive, a person stays in the loop.
What we do not do at all
A shorter list, and more useful than most of what is above.
- We do not buy links. It works until it does not, and the recovery costs more than the gains.
- We do not publish thin location pages. A page that is your home page with the city name swapped is detectable, and it drags down the pages around it.
- We do not guarantee positions. Nobody controls the ranking. Anyone promising position one is quoting a number they cannot deliver.
- We do not write testimonials. Illustrative composites get labelled as composites. Invented client quotes pay off exactly once.
- We do not report only good news. A report where nothing ever falls is not a report. It is a newsletter.
What the monthly report should contain
Reporting is where the whole arrangement is won or lost, and it is the part most providers treat as an afterthought. A useful report fits on one page and contains exactly five things.
- Every keyword you have claimed, with last month beside this month.Not a selection. Not the top ten. All of them, so the ones that went nowhere are as visible as the ones that moved.
- What published, by URL.Not “four pieces of content” — four links you can open.
- What fell, and the current theory why. A theory is allowed to be wrong. Its absence is not.
- What is queued for next month. So you can object before the work happens rather than after.
- One plain sentence on whether the trend justifies the invoice. This is the line that costs providers money to write honestly, which is exactly why it belongs there.
If a report needs a walkthrough call before it makes sense, it was written to be presented rather than read. Those are different documents with different purposes, and only one of them is for you.
Where the money actually goes
For a typical small-business month at the $600–$1,200 mark, the hours split roughly like this. Percentages move with how broken the site was, but the shape holds.
| Activity | Share of the month | Mostly done by |
|---|---|---|
| Writing and editing pages | ~40% | Agent drafts, human edits |
| Research before writing | ~20% | Agent, human picks the angle |
| Technical checks and fixes | ~20% | Agent, human triages |
| Monitoring and diagnosis | ~10% | Agent watches, human explains |
| Reporting and correspondence | ~10% | Human |
The useful thing about seeing it laid out is noticing what is missing: there is no line for meetings, and no line for account management as a separate discipline. In an engagement this size, the person doing the work should also be the person you email. Every layer between those two roles is a layer you are paying for.
How to tell if yours is doing the work
Five questions. Ask your current provider, or ask us. Each should get a direct answer inside a minute, and the ones that do not are the interesting ones.
- Which keyword is each page on my site targeting? (Expect a list.)
- What did you deliberately choose not to target, and why? (Expect a skip list. Its absence is the answer.)
- Which URLs published last month? (Expect URLs, not a count.)
- Which positions fell last month? (If the answer is “none”, nobody is looking.)
- If we stopped tomorrow, what do we keep? (Should be everything.)
Adaeze, from the top of this piece, could not answer any of the five after eleven months. That is not because her agency were crooks. It is because nothing in the arrangement ever required anyone to write the answers down.
The work itself is not glamorous and it is not mysterious. It is a keyword list, a publishing rhythm, a checklist run properly every week, and a report that includes the bad news. Everything else is packaging.