Ask a coding agent to write a meta description and it will happily target a keyword. Ask it why that keyword, and it will tell you the keyword “gets decent search volume.” Ask it for the number, and it will give you one.
That number is almost always invented.
A model cannot know search volume
Search volume is a live, market-specific metric. It changes month to month, varies by country, and is sold by a handful of data vendors. None of that lives in a language model’s weights. When an agent says a keyword does “around 8,000 searches a month,” it is not reading a value — it is producing a plausible-sounding token sequence shaped like a number.
This is not a bug in the model. It is doing exactly what it was trained to do: continue the text convincingly. The problem is that convincing and correct are different properties, and SEO decisions need the second one.
Confident guessing is worse than no answer
If an agent said “I don’t know the volume — you should check a tool,” you would go check a tool. Instead it hands you ~8,000, and that number quietly becomes the basis for a content plan, a page title, an internal-linking decision.
Three weeks later someone asks why the team built a page around a keyword that gets 90 searches a month. The honest answer is: a model rounded a guess to a confident-looking figure and nobody questioned it.
The two bad workarounds
Once you notice the guessing, you reach for a fix. There are two obvious ones, and both have sharp edges.
Let the agent browse. Point it at a browser tool and let it search Google. Now it sees real SERPs — but it sees them as full pages of prose. It reads ranked snippets, ads, related questions, and footer links, then tries to reason a volume estimate out of pixels. You have traded a clean guess for a noisy one, and burned a few thousand tokens of context doing it.
Wire up a raw provider API. SEO data vendors sell real numbers through APIs. But those APIs are built for dashboards: deeply nested responses, hundreds of fields, batch endpoints with their own paging rules. Your agent now needs SDK setup, response mapping, and error handling before it can read a single volume figure.
Both work. Neither is something you want an agent doing inline, mid-task, every time it needs one number.
What an agent actually needs
The fix is not more browsing or more glue code. It is a tool shaped like the thing an agent is good at calling: one command, one predictable answer.
seoctl volume "seo api" --country us{"keyword":"seo api","volume":1900,"source":"clickstream"}That is the whole interaction. The agent asks a precise question, gets a flat JSON object, and moves on. No page to interpret. No envelope to unwrap. No estimate to second-guess.
When the real number is one command away, the agent stops guessing — not because you told it to, but because guessing is now the harder path. The quickstart gets the CLI running; agent setup hands the whole thing to your coding agent.
The rule worth enforcing
Give your agent a real data source, then tell it one thing: SEO numbers come from the tool, never from memory. If the tool fails, surface the error. Do not fall back to an estimate.
An agent that says “I don’t know yet, let me check” is worth far more than one that always has an answer. SEOCTL exists to make checking the easy option.
Try it
Give your agent real SEO data.
One CLI command for keyword volume, SERP, rank, and competitor research. Start with the $7 trial, then scale with monthly credits.