AI for SEO, and What Running It Really Costs

The going rate for a click on AI for SEO is $45.04. That's what my keyword tool reported in July 2026, alongside about 8,100 monthly searches and a difficulty score of 28. Forty-five dollars a click is the kind of number that explains why the ranking pages read the way they do.

Two of the nine organic results my checker returned are Reddit threads, which is the tell that the commercial pages aren't fully satisfying the query.

Direct answer up front. AI for SEO is genuinely useful for three things, which are processing keyword and SERP data at a volume you couldn't do by hand, drafting from research you've already gathered, and finding structural gaps in what already ranks. It's unreliable for judgment calls, and it does nothing at all for the part that most limits a new site. The whole data layer costs less than a dollar a day at the scale a solo operator works at, and I'll break that down to the cent below, because I've never seen anyone else do it.

What The Data Layer Actually Costs

Every guide on this topic lists tool subscriptions. Nobody prices the raw inputs, which is odd, because that's where the interesting economics live.

I pulled these from DataForSEO's own pricing pages on 28 July 2026. Google Organic SERP API runs at $0.6 per 1,000 SERPs on the standard queue, $1.2 per 1,000 on the priority queue, and $2 per 1,000 in live mode. Keyword data through their Google Ads endpoint is $60 per million keywords on the standard queue and $90 per million in live mode, with tasks accepting up to 1,000 keywords each. Their pricing overview notes a minimum payment of $50.

Now put that into a real job. Say you're mapping one niche properly, which for me means pulling a large keyword pool, filtering it, and then checking live search results for the survivors.

Three dollars and eighty-four cents to map a niche. That number surprised me the first time I ran it and it still does.

Spread across the thirty or so articles a niche map supports, the data works out at about 13 cents per article. Add the drafting tokens, which on Claude Sonnet 5's introductory rate of $2 per million input and $10 per million output comes to roughly eleven cents for a research-heavy 2,000-word draft, and one detector scan at about fifteen cents on Originality.ai's $14.95 Pro plan where a credit covers 100 words.

Forty cents an article in hard costs. The real barriers are the $50 minimum deposit and the fact that none of this tells you whether the article was worth writing.

Three Things The Pipeline Taught Me

I've been running this setup long enough to have collected some scars. These are the three that changed how I work.

Keyword expansion drifts toward volume. Ask a keyword API for terms related to a seed and each hop of expansion slides toward whatever adjacent topic has more search demand. Two or three hops out and your list belongs to a different niche entirely, one that's bigger and more competitive than the one you started in. It happens quietly because every individual keyword looks plausible. I lost a couple of weeks to this before I noticed I was planning content for a market I had no business entering. Now I check every expanded list against the seed and throw out anything that drifted, which usually means throwing out the highest-volume half.

Brand-navigation queries look like free wins and are not. You'll find terms with real volume and a low difficulty score where the whole top ten is one company's own pages. It scores as easy because there's no competition in the usual sense. But everybody searching that phrase is trying to reach a specific website, and you are not that website, so ranking second gets you nothing. The difficulty metric can't see intent. I've written pages targeting these and they did exactly what you'd predict, which is nothing.

A soft top ten is the best signal a new site gets. When the results page contains a Reddit thread, a Medium post, or a forum, that's a slot a small site can genuinely take, because those results are there by default rather than by strength. This is the single most reliable thing my scans surface and I now weight it above difficulty scores entirely. Two Reddit results, like the ones sitting in this page's own target query, mean two positions that aren't defended.

None of those three insights came from a tool. They came from running the tool and then being wrong repeatedly.

What Google Says About Optimising For AI

There's a whole emerging industry selling AI Overview optimisation, and the documentation is unusually blunt about it.

Here's Google's page on AI features, loaded 28 July 2026. "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." On eligibility, "a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. There are no additional technical requirements."

And then there's the sentence that ought to end several product categories. "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add."

The same page does note that "all existing SEO fundamentals continue to be worthwhile", which is the part the optimisation vendors quote and the rest of which they tend not to.

I'll caveat this properly. That's the documented position in July 2026 and I have no idea how durable it is. But if someone is selling you a special file format to get cited in AI answers, the vendor's own platform documentation currently says you don't need one.

Where AI Genuinely Helps

Being fair to the technology after being rude about the marketing.

Volume work is the honest win. Reading 120 search results pages and characterising each one is a task that takes a person a full day and a script a few minutes. Same with clustering several thousand keywords by intent, or comparing what six ranking articles cover against each other to find the section nobody wrote. These are jobs where the answer is genuinely mechanical and the only obstacle was tedium.

Drafting from research you already gathered is the second one, covered properly in how AI fits into a writing workflow. The order matters more than the tool.

Structural gap analysis is the third, and it's underrated. Feed a model the actual text of the pages currently ranking and ask what question none of them answer. The output needs heavy filtering because it'll suggest obvious things, but roughly one suggestion in five is a real gap, and one real gap is the difference between an article that ranks and one that joins the pile.

For the repeatable end of this, the scheduling and monitoring parts, SEO automation is a separate discipline with its own tooling. And if you're generating pages from a data set rather than writing them individually, programmatic SEO is a different problem again, with a much sharper policy risk attached to it.

Where It Doesn't Help At All

The judgment calls. All of them.

Whether a keyword is worth targeting. Whether the intent behind a phrase matches what you sell. Whether the gap you found is a gap because nobody's written it or because nobody wants it. Whether a difficulty score of 28 means anything for your specific site. A model will answer all of those confidently and the confidence is not correlated with accuracy.

And then there's the one that took me longest to accept.

I had a site sit at an average position of 47 in Search Console for months. Google's own documentation defines that metric as "the topmost position occupied by a link to your property or page in search results, averaged across all queries in which your property appeared", and notes that "A link must get an impression for its position to be recorded". So the pages were indexed, they were being served, and they were being served somewhere nobody scrolls to.

The first sale on that site was $29.99 and it arrived about two months after launch. I spent most of those two months improving the pages, which was comfortable and pointless, because the pages weren't the problem. Nothing in my tooling was going to fix it either. A keyword API tells you what to write. It has no opinion about whether anyone will ever link to you, and links are most of what separates position 47 from position 7.

A Practical Setup

If I were building this from scratch today, roughly this.

A keyword-data API on pay as you go rather than a monthly seat. The $50 minimum deposit lasts an extremely long time at solo-operator volumes, given a full niche map costs under four dollars.

A live SERP checker, which is the same API, used specifically to count how many results in each top ten are forums, Reddit, Medium, or otherwise soft. That count is the number I actually act on.

A general model with an API key for drafting and for reading ranking pages at volume. Not a specialised SEO writing product. The specialisation in those is mostly prompt scaffolding and prompts are free.

Nothing else, at first. The category is full of $99-a-month platforms that wrap the same two data sources you can hit directly for pocket change, and the wrapper is worth paying for only once your time is genuinely more expensive than the subscription.

Then the uncomfortable part, which is that after all of that you still have to make people aware the site exists. Search doesn't work for a new domain for months, and no amount of data quality shortens that. What the data does is make sure that when the site finally becomes visible, it's visible for something people actually search for. That's worth a lot. It's just not the same as traffic, and I spent a while confusing the two.

For the wider question of what tooling is worth buying across the whole content stack, the AI writing tools breakdown has the subscription math. And if the goal is specifically climbing, getting a site to the top of Google is the harder half of this and it's mostly not about AI at all.