What people
actually type
Google finishes your sentence based on what millions of other people searched for. This tool asks it to finish yours a few dozen times, sorts the answers by intent, and flags the ones you do not want.
Comma separated, up to 12. Put the geography here, never in the keyword. A town in the keyword box returns almost nothing.
| Group | Predictions pulled | What it finds |
|---|---|---|
| Buyer intent | 15 | Ready to spend. Ad keywords and service page targets. |
| Questions | 14 | Still researching. Blog posts and FAQ answers. |
| Comparisons | 8 | Deciding between options. Explainers that win the choice. |
| Local | up to 12 | One request per town. Which places have demand at all. |
| A to Z | 26 | Wide net. Service variations, brands, place names. |
| Prepositions | 13 | Qualified intent. Who it is for and what it is with. |
Type a keyword above. You will get back the phrases Google predicts when someone starts typing it, grouped by what the searcher is probably trying to do.
Why this exists
Most keyword tools hand you a spreadsheet and a volume column and let you assume the hard part is done. The hard part is not collecting the words. It is noticing which of them are lying to you.
This tool pulls the same free data anyone can get, then does the part nobody bothers with. It marks the predictions Google substituted rather than found, drops the searches that came from students and job hunters, and sorts what survives into four lists that each map to a different piece of work. A keyword dump is not a deliverable.
There is a second reason, and it has become the more interesting one. Language models retrieve by comparing the shape of a question against the shape of your content. That comparison is far less forgiving than a search engine, which spends enormous effort working out what you probably meant. Retrieval mostly just measures distance.
Which makes a list of questions real people actually type unusually valuable, and not as keywords to sprinkle into a page. It is the literal phrasing to answer, in headings, in the order people ask it. The same list doubles as a test set: run it against your own site or assistant and count how many come back with a usable answer. That number is the closest thing to a measurable content score you will get for nothing.
Methodology in full
Everything below describes what happens between pressing search and reading a chart. None of it is proprietary and none of it is complicated, but you should know how a number was produced before you make a decision with it.
What happens when you press search?
Your keyword is paired with every term in the groups you checked, one request each. Buyer intent sends 15, questions 14, A to Z 26, prepositions 13, comparisons 8, and Local one per town. Requests go out six at a time. Each reply is trimmed to remove your own keyword and anything already seen, then what is left is grouped, screened, and charted. A full pull of every group is 76 requests and takes a few seconds.
Where does the data come from?
Google's public autocomplete endpoint, live at the moment you press search. It is the same system that drops predictions under the search box as you type, and those predictions come from real searches people perform. The endpoint is undocumented and unsupported, so it can change without notice. Results are cached for seven days.
Why does a longer keyword return less?
Autocomplete completes a prefix, so every word you add makes that prefix rarer. Measured on this tool: "web design" returned 141 phrases, while "web design long island" returned 1. Keep the keyword to two or three words and put any geography in the Local box instead, where it becomes the modifier rather than part of the prefix.
Why do counts stop at exactly 10?
The endpoint returns at most ten predictions per query. Any modifier with real demand behind it saturates there, so ten means "at least ten" rather than a measurement. Those bars are marked. The readable signals are the modifiers that came back under ten, which means demand is genuinely thin for that framing.
What does "substituted by Google" mean?
When the endpoint has nothing for a query it does not return empty, it returns something loosely related instead. Asking for completions of "web design huntington" can hand back "web design card examples", which contains no Huntington at all. Any prediction that does not contain the term that produced it is marked and dropped from the charts, because it is an artifact of that fallback rather than evidence of demand. Every other keyword tool exports these silently.
How are the two charts calculated?
The modifier chart draws two bars. The faint one behind is everything a modifier returned; the solid one in front is what survived screening. The gap between them is noise. The vocabulary chart is term frequency across surviving phrases after discarding words under three characters, common stop words, your own keyword, and the modifier terms. Removing your keyword is the step that matters, or it tops every chart and tells you nothing.
What can this not tell you?
Search volume, competition, or difficulty. A phrase appearing means enough people searched it for Google to predict it, and nothing more precise. Predictions are also shaped by location and, for signed in users, personal history. Treat everything here as candidates worth checking, then price them somewhere with real data before committing.
Everyone reads the same dashboards. Clever marketing is found in the data nobody thought to go and pull.