How to find keywords for a new website with no Search Console data

How to find keywords for a new website, when Search Console has nothing in it yet: pull candidate phrases from three sources that do not need your own traffic, then score each one on five checks and keep only the phrases that pass at least three. The sources are autocomplete, the questions the results page itself asks, and the words your first users already use. The scoring table is at the end of this chapter and it fits on one screen.

Read this first

The moment Search Console has 28 days of impressions, stop using this chapter and switch to striking distance keywords in Search Console. Your own data beats every method below, because it tells you what you already almost rank for. This chapter is for the window before that data exists, which for a new site is roughly the first two months.

One thing this chapter cannot give you is search volume. Volume numbers come from paid tools or from Google Ads with active spend. Everything here proves that a phrase is typed by someone, and none of it proves how many someones. Treat the output as a shortlist to write against, not as a forecast.

Three places to find keywords for a new website

Each source proves a different thing, which is why one alone is not enough to act on. Autocomplete proves the phrase exists in the wild. The results page proves what shape of answer wins. Your own users prove which words the buyer uses before they know your product category by name.

Source 1 — autocomplete tells you what people start typing

Autocomplete is the cheapest real signal available to a new site. If a phrase produces suggestions, people finish that phrase often enough for Google to have learned it. If it produces nothing, the phrase is something you invented.

curl -s "https://suggestqueries.google.com/complete/search?client=firefox&hl=en&q=indexnow" \
  | python3 -c 'import json,sys; print("\n".join(json.load(sys.stdin)[1]))'

Run it on every phrase you are considering, including the ones you are sure about. Planning this batch of chapters on 2026-08-17, we ran 46 English candidate phrases through it and 26 returned no suggestions at all — among them how to set up indexnow and which ai crawlers should i allow, both of which read perfectly well in a title.

The list also moves. We ran the same 46 phrases twice on that one day, and ai crawler user agent list came back empty the first time and populated the second. So treat a single empty result as weak evidence and a repeated one as strong.

Two cautions. This endpoint is not a documented API, so it can change or start refusing requests without notice; the same list is visible by typing into the search box. And the hl parameter matters more than people expect: the same phrase can be rich in one language setting and empty in another.

Source 2 — the results page asks the next question for you

Search your candidate phrase and read the page instead of the rankings. People Also Ask boxes and the related searches strip at the bottom are Google telling you, in real users' words, what the follow-up questions are. Those follow-ups are usually longer, less contested, and easier for one page to answer completely.

Copy those questions down verbatim. Do not tidy the grammar. A phrasing that looks clumsy is often the phrasing that gets typed, and it makes a good H2 exactly because it matches what someone asked.

Source 3 — the words your first users already use

Support messages, sales calls, and forum threads carry vocabulary you would never guess. This source is small and slow, and it is the only one of the three that your competitors cannot copy. Read your own inbox for the nouns people use for the problem, then check those nouns against source 1.

If you have no users yet, the substitute is a forum where your buyers argue. Read threads, not titles. The phrase you want is usually in the third reply, where someone restates the problem in plain words.

The scoring table

Score each candidate out of five. Three or more, write the page. Two or fewer, drop the phrase and take the longer version of it instead.

Two of these are judgement calls and that is deliberate. Reading a results page for intent is something a person does in ten seconds and a script does badly. If you find yourself arguing to give a phrase its third point, that argument is the answer.

What to do when a phrase scores two

Do not force it. Take the phrase and extend it into the question someone would actually ask, then re-score the longer version. A phrase that fails on competition almost always passes once it becomes a full question, because the long form has fewer pages written against it.

• ai visibility fails, so score how to check if chatgpt can crawl my site instead.
• seo for ai fails, so score why is my page cited by perplexity but not chatgpt instead.
• Do not keep a two-point phrase because it is the one you wanted.
• Do not stack three near-identical phrases onto one page hoping one lands.

Three ways this goes wrong

All three are ways of quietly turning the method off while still going through the motions. The first is the one that catches experienced people, because it feels like diligence.

1. Scoring the phrase you already wrote the page for. Run the checks before drafting, or you will find reasons to pass it.
2. Trusting autocomplete as a volume signal. It shows that a phrase is typed. It says nothing about how often, and a rare phrase can suggest just as readily as a common one.
3. Picking a phrase in the wrong language setting. A phrase can be strong under hl=en and empty under another, and the market you are writing for decides which one counts.

What this method will not do

It will not rank you. It filters out phrases that were never going to work, which is a different and smaller job. It also cannot tell you whether the page you write will be good enough to hold a slot once it gets there — what makes content citable by AI is the chapter for that half.

And it will not stay accurate. Autocomplete lists change, results pages get restructured, and a phrase that scored four in August can score two by spring. Re-score anything you are about to invest a week in. Keeping that check honest across a whole site is the part QueryWin is being built to carry.

Common questions

How many keywords should a new site start with?

Fewer than feels comfortable. One phrase per page, and enough pages that you learn something from Search Console within two months. The point of working out how to find keywords for a new website is to stop guessing, not to build a longer list — ten scored phrases beat sixty guessed ones.

What if every phrase in my niche scores two?

Then the niche is contested at the head and your entry is further down the tail. Go to source 3, find how your buyers describe the problem before they know the product category exists, and score those words.

Can I use allintitle without knowing volume?

Partly. The ratio needs both numbers, but the raw allintitle count alone is still useful: a phrase with thousands of exact-match titles is contested regardless of what the volume turns out to be.

Does this work for a language other than English?

The method does; the thresholds may not. We have run these checks in English and Chinese and found the same phrase can be strong in one and produce nothing in the other, which means the two lists have to be built separately rather than translated.