Choosing Keywords for an AI-Written Article: A Beginner's Field Guide

The keywords field in Smart-Copy.ai accepts at most ten phrases, and that constraint is the whole point: with ten slots, what you put in them matters more than anything you do to the draft afterwards. The generator honours those phrases exactly as promised — primary phrase into the heading or opening paragraph, the rest woven through the body, no stuffing. What it will not do is tell you that you picked the wrong ten.

Where candidate keywords actually come from

Keyword research starts by collecting, not choosing. At this stage you want 30 to 60 candidates, from which five to eight will survive. Four sources are enough to build that pool, and three of them cost nothing.

1. Google Search Console — the only data about your site

The Performance report lists the queries your site already appears for, with impressions, clicks and average position, going back up to 16 months. It beats every third-party estimate for one reason: it describes your domain, not the market.

The useful filter is average position between 8 and 25 with meaningful impressions and poor click-through. Those are queries where Google already considers you partly relevant but not relevant enough for the top three. An article aimed at one of them has a much shorter path to page one than a piece targeting a phrase you have no history on at all. Read average position carefully, though — it is an impression-weighted average across devices, countries and result types, not a fixed rank.

2. Google Keyword Planner — useful, with a caveat

Keyword Planner suggests related phrases and estimates volume. The catch: a Google Ads account without an active, funded campaign sees volumes bucketed into broad logarithmic ranges. A bucket like "1K–10K" spans an order of magnitude — the phrase might get 1,200 searches a month or 9,000, and those two numbers justify completely different editorial decisions.

So use the Planner for idea generation and grouping, not for prioritisation. Prioritise with Search Console, where the numbers are real, or with a paid tool.

3. The search results page itself

Autocomplete, the "People also ask" box and the related searches at the bottom of the page list phrases Google itself considers adjacent to your topic. For how-to content, those questions convert directly into H2 and H3 headings, and answering them closes out the reader's intent rather than sending them back to search.

4. Paid tools

Ahrefs, Semrush and their equivalents add two things free sources lack: a specific volume figure instead of a range, and a measure of ranking difficulty derived from the link profiles of the current top ten. For one article a month that is overkill; for planning twenty or thirty, it is the difference between a plan and a guess.

Picking the primary phrase and the long-tail set

One article gets one primary phrase plus four to seven supporting ones. Not ten equals — one phrase the entire piece is built around.

The primary phrase describes the whole article, not a section of it. The test: if someone typed that phrase into Google and landed on your piece, would they get a complete answer? If yes, it is your primary. If they would get one paragraph, it is either a supporting phrase or a separate article.

Supporting phrases are usually long-tail: longer, more specific, lower volume, less contested. "Keyword research" is broader than "how to choose keywords for a blog post", but the second tells you far more about what the reader expects to find — and a page that matches an expectation ranks better than one that matches a string.

Check intent before you commit

The most common beginner mistake is picking a high-volume phrase whose intent does not match the article you plan to write. A transactional phrase ("best tool for…") calls for a product page or a comparison; an informational one ("how to…", "what is…") calls for a guide. One article will not serve both.

Verification takes thirty seconds: search the phrase and look at what type of page occupies the top ten. All e-commerce listings while you are writing a tutorial? Change the phrase. The results page is Google's verdict on the right answer format, and a single article will not overturn it.

Drop phrases that cannibalise each other

If two phrases on your list return nearly the same set of URLs in the top ten, Google treats them as one query. Do not build two articles around them. Put both in the same piece and avoid the situation where your own two pages compete for one position.

What Smart-Copy.ai does with the phrases you supply

Phrases go into the keywords field of an order, from the dashboard or via POST /v1/texts in the API. How they are handled should shape how you choose them.

• The primary phrase lands in the heading or the first paragraph. That is a rule enforced during generation, not a suggestion offered afterwards — so the first phrase on your list should be one that reads well in a title.
• Remaining phrases are placed inline, inflected to fit the sentence. They are not pasted in as rigid strings, which is why you can submit them in base form.
• Synonyms and long-tail variants are added on top. You do not need to spend your ten slots on morphological variants of one concept; spend them on ten different concepts.
• Internal links are a separate field. links takes up to five URLs with optional anchor text, placed and validated independently of the keyword rules — the mechanics are covered in Automated Internal Linking: Anchor, Position, Validation.

For the full pipeline, see How SEO Optimization Actually Works Inside Smart-Copy.ai. And if you are weighing this against recommendation-based tools, Surfer SEO Alternative? What Smart-Copy.ai Does Differently sets out the difference between advice you implement and rules applied at generation time.

Keyword density: the number not worth counting

"How many times should I repeat the phrase" has been asked in every SEO guide for two decades, and the answer is uncomfortable: there is no optimal keyword density. Google's John Mueller has said publicly, for years, that no concept of optimal keyword density exists at Google — while adding that mentioning the topic explicitly does still matter.

The other end of the same scale is documented, though. Keyword stuffing — loading a page with phrases to manipulate rankings — is named outright as a spam practice in Google Search Essentials. There is no percentage threshold at which text "becomes" spam; there is a judgement about whether the repetition serves the reader.

What that means in practice:

Instead of	Do this
Targeting 2% density	Confirming the phrase appears in the title and the opening
Repeating the phrase in every H2	Using variants and "People also ask" questions as headings
Submitting ten variants of one phrase	Submitting five to eight distinct concepts from the topic
Counting occurrences after writing	Reading the draft aloud and cutting sentences that sound forced

The last row is the most effective and the least used. A phrase wedged in against the grain of the sentence is audible on a read-through long before any tool counts it.

Checklist before you submit the order

1. One primary phrase covering the whole topic, first on the list.
2. Four to seven supporting phrases, at least two of them long-tail.
3. All phrases sharing one intent — informational or transactional, not mixed.
4. No two phrases that return the same top ten.
5. Base forms only; inflection is the generator's job.
6. Audience and context written into guidelines, not appended to the keyword list.

That last point is the one people get wrong most often. The keywords field names the phrases; it is not where you describe the article. Editorial requirements — tone, audience, what must be covered — belong in guidelines, and your own reference material goes in source_urls. Keeping the three separate produces a visibly better draft than cramming everything into one box.

Good keywords will not rescue thin content. But bad ones will waste good content: the article ends up ranking for queries nobody types, or competing with a post you published last year. Half an hour in Search Console before you write pays back faster than any optimisation you attempt afterwards.