How to Determine Search Intent for Keywords (Free Workflow)

You picked a keyword. You are about to brief a writer or open a blank doc. Stop. The first question is not "how do I rank for this," it is "what does the searcher actually want?" Get that wrong and the article underperforms even if every other SEO move is perfect.

This guide shows how to determine search intent for keywords in five minutes, using a free LLM and the live Google results page. No SEMrush. No Ahrefs. No paid intent score. You hand the model a structured prompt, it reads the top five titles, and you walk away with the angle, format, and three H2s for your next article.

Why search intent matters more than search volume

Volume tells you how many people search. Intent tells you what they came to do. A 50,000-volume keyword you mismatch loses to a 500-volume keyword you nail.

Here is the thing. Google has spent the last decade tuning rankings around how well a page resolves the searcher's job. The four-layer Ranking Signal Stack that drives Google SEO today puts query interpretation above almost every on-page factor. Intent is the layer most small sites get wrong.

Volume without intent costs you twice. Once on the writing — wrong format, wrong angle, wrong length. Twice on the click — visitors bounce, dwell time tanks, the page slides down the SERP within weeks.

Action: before you brief any article, decide the intent. The rest of this guide shows you how to decide it in five minutes without paid tools.

The four intent types and what they mean for content format

Search intent type diagram: informational, navigational, commercial, transactional cards on dark slate

Every keyword falls into one of four intent buckets. Each bucket maps to a page format Google already ranks for that query. Your job is to match.

Informational — the searcher wants to learn something. Format: long-form guide, explainer, list. Example query: "what is search intent."
Navigational — the searcher is looking for a specific brand or page. Format: branded landing page or homepage. Example: "varynforge login."
Commercial — the searcher is comparing options before buying. Format: comparison, alternatives, "best X for Y" list. Example: "best keyword research tool."
Transactional — the searcher is ready to buy or sign up. Format: product page, pricing, checkout flow. Example: "buy ahrefs subscription."

The buckets are not new. The skill is reading them off a live SERP in seconds. If the top 10 results are all "best X" lists, the keyword is commercial — even if your gut said informational. The SERP is the only ground truth.

For a deeper score across multiple axes when the SERP is mixed, see the four-axis vector framework for search intent. For this five-minute workflow, the four-bucket call is enough.

How the "Decode search intent" prompt works

The prompt does three things. Classifies the intent. Reads the SERP for you. Hands you an angle that beats the top three results.

Inputs are minimal: your niche in one line, the target keyword, and the top five page titles from a Google search. No API key, no scraping, no paid SERP data.

Output is structured: intent class, evidence from the SERP, dominant content format, the angle a smaller site can win with, and three suggested H2s. Paste it straight into a brief.

Here is the prompt verbatim. Copy it.

"You are an SEO analyst. Niche: [your niche, one sentence]. Target keyword: [keyword]. Top 5 Google results for this keyword (titles only): 1) [title] 2) [title] 3) [title] 4) [title] 5) [title]. Do four things. First, classify the intent as informational, navigational, commercial, or transactional, and quote the SERP evidence. Second, name the dominant content format Google is already ranking. Third, propose one angle a smaller site can win with — must be different from the top three results. Fourth, list three H2 headings that match the intent and the angle. Do not invent search volumes. Do not recommend keyword tools."

The constraints matter. "Do not invent search volumes" stops the model from hallucinating numbers — a documented LLM failure mode. "Do not recommend keyword tools" keeps the output focused on intent instead of pivoting to a generic SEO tool list.

Walkthrough: paste any keyword, get back the angle to write

LLM chat mockup showing a Decode search intent prompt with structured four-part output on dark background

Take a real keyword. We will use "keyword research for beginners" — informational on the surface, but the SERP tells a more specific story.

Gather the inputs. Search the keyword in an incognito window so personalization is off. Copy the titles of the top five organic results. Skip ads, skip People Also Ask, skip the AI Overview.
Write the niche line. One sentence. "I run a content marketing blog for small ecommerce founders" is enough. The model uses this to tailor the angle so the output is not generic.
Paste into ChatGPT, Claude, or Gemini. Any frontier model works. Free tiers handle this fine — the prompt is short, the output is short.
Read the output critically. Cross-check by skimming one or two of the top results. If the intent does not match what those pages actually deliver, rerun with the SERP descriptions added — titles alone occasionally mislead on commercial-vs-informational edge cases.
Turn it into a brief. Take the format, the angle, and the three H2s straight into your outline. The five-minute spend just saved you a 1,500-word rewrite.

The screenshot below shows this running inside Claude with the keyword "keyword research for beginners" and an ecommerce-blog niche. The model classifies the intent as informational with a commercial undercurrent, calls out that the top results are tutorials with embedded tool recommendations, and proposes an angle aimed at solo founders who do not want to pick a tool yet.

The angle the model surfaced — "the workflow you run before you pay for any tool" — is the kind of thing a smaller site can rank for. It does not try to outgun Ahrefs and Semrush on the head term. It carves a slice and owns it. If your output is generic, your niche line is doing too little work. Rewrite it with one specific reader pain. Rerun.

How to use this before briefing a writer or outlining an article

The prompt is a brief input, not a brief replacement. Here is the workflow we run on every article in the editorial calendar.

Run the prompt. Save the output as a comment in the brief doc. Two minutes.
Confirm the format. If the model says "long-form guide" but your slot is a 600-word listicle, change one or change the other. Do not fight the SERP.
Take the angle into the title. If the angle says "for solo founders," the title should say "for solo founders." Vague titles match vague intent.
Use the suggested H2s as a starter outline. Cross-reference against the keyword-type decision tree to make sure the angle is one a smaller domain can actually convert.
Hand to the writer with the SERP titles attached. Writers who can see the SERP for themselves write better drafts than writers given an abstract intent label.

The compounding payoff: once you do this for ten keywords in a cluster, you start spotting which clusters are worth writing at all. Some clusters are dominated by transactional commerce pages no informational article will dent. Skip those. Spend the time on the clusters where the SERP has weak informational coverage.

Do this in 5 minutes: how to determine search intent for keywords on your backlog

Five-step search intent workflow: Pick, Search, Prompt, Read, Brief — connected by gold lines on dark slate

Right now, before you close this tab.

Pick one keyword from your editorial backlog. The one you were about to brief next.
Search it in an incognito window. Copy the top five organic titles.
Open Claude or ChatGPT. Paste the prompt above. Fill in your niche, the keyword, and the titles.
Read the output. Note the intent class, the format, and the angle.
Update the brief. Adjust the title and outline to match.

Total elapsed: under five minutes. You just made every downstream hour — research, drafting, editing — point at the right target.

Weaknesses and drawbacks of the free-LLM workflow

The prompt is a sharp tool for one keyword. Be honest about its limits.

Where the free LLM falls short

No live SERP. Every call is a guess from training data, not what ranks today.
Training cutoffs hide AI Overviews and new snippet formats from the model.
One keyword at a time. Fifty keywords means fifty round-trips — exhausting at backlog scale.
No ranking-page analysis. It does not read the top results, their depth, or formats.
Uniform confidence on obvious and mixed intents — you catch the difference.

What you reach for next

When the manual workflow stops scaling, the upgrade path is live SERP data for every keyword, a confidence score on each intent label, and a map of covered versus uncovered intent. Earn the prompt first; reach for live data when volume forces your hand.

How VarynForge fits in

Running the prompt once is easy. Running it across thirty keywords without losing the outputs is the harder problem. VarynForge keeps the "Decode search intent" prompt in a saved library alongside six others — niche research, content-gap audit, page analysis, outline, titles, brief — and persists every output under the keyword it was run against. Create a free VarynForge project to use the prompt library and the free brief generator (10 briefs per 24 hours, no credit card).

Frequently Asked Questions

How is this different from a paid intent score in SEMrush or Ahrefs?

Paid tools assign an intent label per keyword based on aggregate SERP analysis. The prompt does the same job for one keyword at a time, free, and shows its work — you see the evidence the model used. For an early-stage site running a few briefs a week, the prompt is enough. Once you are scoring hundreds of keywords a month, paid tools save time.

Can I trust the LLM to read the SERP correctly?

For the four-bucket classification, yes — the model is reading title patterns, which is a task LLMs are reliable at. For volume, difficulty, or anything numeric, no. That is why the prompt explicitly tells the model not to invent search volumes. Treat the output as a fast first read; spot-check by clicking one or two of the top results.

Which model should I use — ChatGPT, Claude, or Gemini?

Any of the three work. We use Claude for the cleaner structured output, but ChatGPT 4-class and Gemini 2.5 produce comparable classifications. Free tiers are fine — the prompt fits well inside any free-tier message budget.

What if the top five results have completely different formats?

That is a mixed-intent SERP and a useful signal: Google has not converged on one format yet, which means there is room. Add the SERP descriptions to the prompt input and ask the model to call out the split. Pick the underserved format if you can match it credibly.

Does this work for non-English keywords?

Yes. Frontier models handle non-English SERP titles well. Write the niche line in the same language as the keyword to keep the model aligned. The output is more reliable in the language you fed it.

How often should I rerun the analysis on the same keyword?

Rerun when the SERP shifts — usually every 6 to 12 months for stable queries, sooner after a major Google update. The format Google rewards changes. The prompt takes two minutes; the cost of running it on a refresh cycle is trivial.

Sources

Conclusion

Search intent is the cheapest leverage in SEO. Five minutes with a prompt and a live SERP saves you from writing the wrong article. The four buckets — informational, navigational, commercial, transactional — map cleanly to a page format Google already rewards. Read the SERP, classify the intent, pick the angle, brief the writer.

Once you have the intent nailed for one keyword, scale the same five steps across your backlog. Calibrate your free Google keyword tools before committing to any paid plan, and review the full seed-to-target pipeline when you are ready to plan a cluster instead of one article.