Most products sold as an "AI SEO agent" are writers. They pick a keyword, produce an article and call it strategy. That is one job out of many, and it is the job least connected to what your site is already doing in search.
An SEO agent worth paying for starts from your own numbers: the searches Google already shows you for, the pages that get shown but not clicked, the pages losing traffic, what Google says about indexing, and which search visitors actually become customers. This page lists what such an agent should read, what it can honestly conclude, what it must never claim, and how to test one in a trial.
If you are choosing a whole company operator rather than one marketing tool, start with the best AI business builder: twelve tests (read 2026-11-02).
What it should read before it writes anything
Google Search Console is the only source of what Google itself recorded for your site: the queries you appeared for, the pages shown, clicks, impressions, click-through rate and the average position Google reports, kept for sixteen months. It runs about two days behind, and it withholds rare queries for privacy, so query rows never add up to the site total. An agent that stores the history day by day can compare periods; one that reads a dashboard once cannot.
Google Analytics answers the question Search Console cannot: what search visitors did after they arrived. Sessions from organic search, by landing page, and the key events and revenue that followed, when the property reports them. Without it, an agent optimises for clicks that may never become customers.
Google Ads Keyword Planner adds demand you do not have yet: average monthly searches for phrases close to what you sell, and how many advertisers bid on them. That competition figure is about paid search. It is not organic difficulty, and an agent that calls it difficulty is guessing.
Your own site is the fourth source. A crawl that reads titles, meta descriptions, main headings, canonical tags, redirects, server errors, thin pages and the sitemap finds problems no report shows until traffic is already lost.
Six opportunities that data can actually prove
Each one comes with the numbers behind it. If an agent cannot show you those numbers, it is not reading your data.
Shown often, clicked rarely
A page with many impressions whose click-through rate is far below your own site's typical rate at the same position. The fix is usually the title and description, and the gain can be estimated from your own numbers.
Close to the first results
A search you already appear for at an average position between about 4 and 20. Making the right page answer it fully, and linking to it from related pages, is cheaper than ranking for something new.
Losing traffic faster than the site
A page whose clicks or impressions fell by 30% or more against the four weeks before, clearly more than the whole site changed. Find the cause first: a changed title, an indexing problem or a newer page taking its searches.
Two pages competing for one search
Impressions for one query split between pages. A signal of overlapping intent, never a certainty, and worth deciding which page should own it.
Searched for, with no page about it
Real impressions for a question no page on your site is about. Demand is proven by Google's own figures, so the new page is not a guess.
A problem Google names
URL Inspection or the sitemap report saying a page is not indexed, blocked, redirected or duplicated. These come before any content work, because a page Google does not index cannot rank.
What it must never claim
A rank it measured by scraping results pages. Search Console's average position is Google's own figure; anything else is an estimate of where a page appeared for someone, somewhere.
Keyword difficulty from paid-search competition, invented search volumes, or a forecast presented as a result. When the data is missing, the honest answer is "not reported", not zero.
That a change caused an outcome. A month later the best an agent can say is that the page moved more than the site did over the same window, with what else changed in that window listed beside it.
What each kind of product reads
| Product shape | Reads your Search Console | Acts on the site | Measures the result |
|---|---|---|---|
| Rank tracker or SEO suite | Often, alongside third-party estimates | No: it reports | Rank movement, by its own measurement |
| AI article writer | Rarely | Publishes articles only | Usually not |
| AI SEO agent inside a company operator | Yes, stored as returned | Titles, links, technical fixes and new pages, through the engineering and content work the company already runs | Your own clicks and conversions against the whole site |
Approval, proof and not repeating itself
Changing a live site should follow a rule you set. Ask about every change, let small fixes such as titles and links run on their own while a rewrite still asks, or let it work on its own within your budget. The approval should show the exact change, not a summary of it.
Done should need proof the platform can see: a published article, a completed engineering task, or a crawl that saw the page change. An agent that marks its own work done without proof will report progress that never happened.
Every opportunity should be remembered with its evidence, confirmed on a second look before it becomes work, dropped when the data no longer supports it, and kept quiet for a while after you dismiss it. Otherwise the same suggestion comes back every week.
Buyer checklist: verify in a trial
Connect Search Console, wait a day, then check:
Your own numbers
The agent quotes queries, pages and clicks you can find in your own Search Console, for the same dates.
Evidence per item
Each suggestion names the rule it applied and the figures behind it.
Unknown is unknown
A metric your Analytics does not report shows as not reported, not as zero.
Competition is labelled
Keyword Planner competition is called paid-search competition.
Your approval rule holds
A change to a live page waits for you when you said it should.
Proof before done
Marking work done fails until something the platform can see has changed.
A measured result
A month after a change, you get a before and after against the whole site, or an honest "too little data".
Where Empyre fits
Empyre's CMO runs SEO this way for every company it builds or runs. The owner connects Google Search Console and Google Analytics in Company settings; the platform stores what Google returns, day by day, and crawls the live site on its own. Keyword Planner ideas come from the owner's Google Ads account when one is connected, and otherwise through Empyre's own Google Ads account, from the company's site, its description and the searches it already appears for. Empyre never creates, changes or pays for anything in an Ads account for this.
From that data the CMO keeps a backlog where every item carries its evidence and rule, works it through the engineering goals and articles the company already runs, under the approval rule the owner chose, and measures each change after 28 days. The owner sees it in Insights on the dashboard: clicks, impressions, click-through rate and Google's average position against the period before, the top searches and pages, visits from search, and what the CMO is doing.
By design it does not track ranks, buy links, scrape results pages or use third-party keyword databases.
Common questions
Do I need a Google Ads account?
Not with Empyre. Keyword Planner runs through Empyre's own Google Ads account when you connect none; connecting yours is optional.
Will it change my site without asking?
Only if you allow it. The default asks before any change to the live site; new articles follow your usual publishing setting.
Does it show my ranking?
It shows Google's own average position from Search Console, labelled as Google's. It never claims a rank it measured itself.
What if my site has almost no traffic yet?
The crawl, indexing checks and Keyword Planner still work. Opportunities that need clicks wait until Google reports enough of them, and results below a threshold are reported as inconclusive.
Try Empyre free for 3 days
Describe a business in plain words and watch eight AI agents build and deploy it. Starter is free for the first 3 days.
Related
Last updated 2026-11-02. Competitor descriptions reflect each product's publicly documented capabilities at that date; they change often, so check the source before relying on a detail.