SERP Intent Classification With AI: Building a Repeatable Query-to-Page Map

Intent classification is usually done by eyeballing a keyword and picking informational, commercial, transactional, or navigational. That label then drives page type, word count, and internal linking. The failure mode is obvious: the label reflects the SEO's assumptions, not what Google actually ranks. A better process reads the SERP as evidence and derives the page format from the result mix, then uses an AI layer only to summarize patterns across many queries.

Read the SERP before labeling anything

Start with the features. A query dominated by shopping carousels, price filters, and product detail pages is transactional regardless of how it reads. A query with a featured snippet, People Also Ask, and long-form guides is informational. A query with a local pack is local, and no amount of blog content will displace map results.

Record the top ten organic URLs with their page type: product, category, comparison, how-to, documentation, forum, video, or news. The distribution is the classification. If seven of ten results are category pages, the intent is commercial browsing, and a 3,000-word essay will not rank no matter how thorough it is.

This manual pass is worth doing for the twenty queries that matter most. For the long tail, automate the collection and let a model label the pattern, with the manual set as ground truth.

What to automate and what to keep human

Collect, per query: SERP features present, top-ten URLs, titles, page types, and word counts. The collection step is deterministic and should be scripted against a SERP API rather than scraped from a browser session, because layouts change and rate limits bite.

Classification is where models help. A prompt that receives the feature list and the top-ten page-type distribution can return a structured label with a confidence score and a short justification. Keep the output schema rigid — {intent, page_type, confidence, evidence[]} — so results can be aggregated. Do not let the model assign intent from the keyword alone; that reproduces the guesswork you are trying to eliminate.

Keep a human-reviewed sample of at least fifty queries. When the model disagrees with the reviewed label, inspect the SERP yourself. The disagreements cluster around hybrid queries (for example, "best X for Y" where the SERP mixes listicles and product pages), and those are exactly the cases where a wrong call wastes the most production time.

Turn intent into a page contract

The useful output of classification is not a label in a spreadsheet. It is a page contract: format, required sections, target entities, internal links, and the conversion element.

For an informational intent with a snippet, the contract is a direct answer in the first 60 words, then supporting detail, then related questions. For commercial investigation, it is a comparison table with pricing, constraints, and a clear recommendation. For transactional, it is a product page with availability and trust signals. Writing the contract before drafting prevents the common mismatch where a category query receives a thought-leadership article.

Validating the map against reality

Classification is a hypothesis. Check it against Search Console after 28 days: query-level impressions, average position, and the page that actually ranks. If a page earns impressions for queries classified as a different intent, the original label was wrong or the SERP is shifting.

Cluster the queries that a page earns and compare them with the intent you targeted. Two outcomes matter. If the page ranks for adjacent informational queries but not the commercial target, split the page rather than stuffing both intents into one URL. If the page earns nothing, the SERP may be owned by a feature — a shopping unit, a knowledge panel, a video carousel — that organic content cannot win.

SERPs move. A comparison-heavy SERP can shift to product pages after a core update, and AI Overviews can absorb the informational layer entirely. Re-run classification monthly for the top hundred queries and quarterly for the rest. Store the history: a query that flips intent twice in six months is unstable, and unstable queries deserve a page that can serve both formats or a decision to skip them.

The rule that keeps this honest: classify from evidence, write a contract from the classification, and verify from Search Console. The model sits in the middle as a summarizer, not as the source of truth.

Instrumenting the classifier so it can be audited

A classifier without an audit trail becomes an opinion with a spreadsheet attached. Log, per query: the timestamp, location, device, SERP feature set, the ordered top-ten URLs, the model's label, its confidence, and the exact prompt version. Store the raw response alongside the parsed label so a later disagreement can be traced to a parsing bug rather than a judgment call.

Version the prompt like code. A one-word change from "classify the intent" to "classify the dominant intent" shifts labels on hybrid queries, and without a version stamp you cannot tell whether a metric moved because the SERP changed or because the prompt did.

Track disagreement rate over time. If human reviewers override 20% of labels in month one and 8% in month three, the taxonomy is converging. If the rate climbs after a core update, the SERP shifted under a stable model, which is exactly the signal worth investigating.

Handling the ambiguous middle

Roughly a fifth of queries will not fit one label cleanly. Three patterns recur.

Two-intent SERPs. Half the results are listicles and half are product pages. Rather than forcing a choice, record both shares and decide whether to target the dominant intent or build a hybrid page with a comparison table above the fold and buying guidance below.

Brand-plus-modifier queries. "Notion AI pricing" returns the vendor's page, comparison articles, and Reddit threads. The intent is commercial, but the domain you can win depends on whether the vendor page is thin. Classify as commercial and check whether the top result is beatable before commissioning anything.

Freshness-driven queries. Results dominated by posts from the last 30 days signal a news or update intent that a static guide cannot hold. Mark these as time-sensitive and route them to a different calendar rather than the evergreen queue.

Store these as attributes, not labels. A page contract can then say "commercial, 60% comparison results, vendor page weak" and the writer knows what to build.

Calibrating confidence against outcomes

Confidence scores are only useful if they predict something. Bucket labels by confidence and compare the 28-day outcome: did the page rank for the target query cluster, or did it earn impressions elsewhere? If low-confidence labels produce most of the misses, route them to human review before production. If high-confidence labels also miss, the SERP is feature-dominated and content investment should stop regardless of label quality.

This calibration is what turns a classification exercise into a prioritization system. The label decides the page format; the confidence and the outcome decide whether the page is worth writing at all.