跳至正文
关键词研究 · Google SEO

How to Choose Keywords for a New Website: From Raw List to Page Plan

// / / 光算科技

For a new website, choose keywords by connecting a real audience need to a page you can deliver. Start with products, services and customer questions, preserve the research conditions, inspect representative results, and map related queries to existing or planned pages. Use keyword difficulty to organize investigation—not as a universal “KD below 10” admission rule.

A useful keyword list ends with decisions: which page to improve, what information it needs, who supplies it and why the work is worth doing. A large export of low-scoring phrases does not establish demand, readiness or a ranking timetable. This guide shows how to turn an initial list into that working plan without inventing missing data.

Start with what the new site can actually offer

Write down the intended market and language, actual products or services, customer tasks, available source material and the actions the site can support. Include essential product and service pages even when their broad terms are competitive. A necessary catalogue page and a near-term search opportunity are different planning decisions.

“No authority” is also an imprecise starting label. A new site may have few visible links or little search data, while its team still has useful product knowledge and documentation. A third-party domain score is not a Google qualification threshold. Assess the work required rather than assuming the site must publish a fixed number of informational posts before it can describe its offer.

Seed ideas can come from product names, sales questions, support language, internal site search where available, and relevant public discussions. Keep their origins distinct. A question raised by one buyer establishes an observed question in that context; it does not establish a monthly market volume.

Keep a raw list and a decision list

Preserve the original phrase. Use separate columns for cleanup and grouping so that a mistaken normalization can be reversed. Remove accidental surrounding spaces and duplicate imports, but do not silently remove model suffixes, negations, sizes, versions or geographic restrictions. Those details may change the answer.

Record groupUseful fieldsWhy they matter
OriginRecord ID, raw phrase, source and collection dateDistinguish an observed customer question from a tool suggestion or an invented idea
Research conditionsLanguage, country/database, tool, metric name and data periodKeep observations comparable and preserve unavailable values
Reader taskNeed, inspected URLs, page types and unanswered conditionsExplain the grouping rather than relying only on an intent label
Page decisionExisting or proposed URL, evidence needed, owner, next action and statusTurn the list into work someone can complete

Adapt the columns to the project; you do not need to buy every metric before beginning. An unqueried value should remain “not queried,” an unavailable value “unavailable,” and a failed request a failure. None should be converted to zero merely to make sorting easier.

If an AI assistant proposes phrases or clusters, label them as generated suggestions until independently checked. It can help organize wording, but cannot supply genuine customer records, measured demand or a product capability your team has not verified.

Expand the list without treating suggestions as evidence of volume

Use a relevant seed and the intended country database in your chosen tool, then inspect variations, questions and related topics. The Semrush research workflow explains how reports can support this process. Exclude clearly unrelated products or places, while preserving potentially useful support and learning queries rather than filtering everything by immediate buying intent.

Google's autocomplete explanation says predictions consider language, location, trending interest and search history; word and phrase predictions can also reflect patterns on the web. Save the conditions you can confirm. A suggestion is a research lead, not a monthly-volume ranking, and its absence does not prove that nobody searches for the phrase.

A new site may have little Search Console data. That is a reason to mark the site's observation as limited, not to manufacture a baseline. When data becomes available, use the performance report with consistent page, query, country, device and date conditions. Privacy filtering, row limits and canonical attribution affect what you see; a missing query row is not proof of zero demand.

Use difficulty after relevance, not instead of it

Ahrefs defines KD as an estimate of reaching Google's top 10 organic results, based on the referring domains of the ranking pages. It explicitly does not include on-page SEO factors. The score cannot tell you whether a query wants your product, whether you have sufficient evidence or whether your page will answer the task.

Use one consistent tool and market to organize a shortlist. Keep other tools' scores separately rather than averaging them. General KD and a metric calculated for a specified domain or URL assess different objects; the keyword difficulty guide explains those distinctions. A score is neither your success probability nor a count of links to purchase.

Volume, difficulty and advertising bids answer different questions. A high CPC does not establish your profit margin, and a low estimate does not make a specific support need worthless. Do not require every candidate to pass fixed KD, volume and CPC cutoffs. If you use a filter to reduce an export for inspection, record it as your working choice and review what it excluded.

Read the actual results before assigning a page type

For important candidates, record the full query, date, language, intended market and any location or device conditions you can control. Separate advertisements and other result features from ordinary web pages, then open representative pages. An API list or a title snippet does not establish that you have reviewed the complete current results experience.

Inspect the task, answer and supporting material. A forum may provide valuable experience; an old manual may still be the applicable version; a short manufacturer page may answer a specification question better than a long generic guide. These observations should replace rules such as “Reddit means weak competition” or “a page older than two years is easy to beat.”

Describe a gap narrowly: “the inspected page does not state the condition for this model” is different from “competitors have no coverage.” Continue into linked manuals or specifications when relevant. If access fails or you only have a fragment, retain that limitation rather than treating the page as poor content.

Use search intent to identify the actual next step. A buyer browsing options may need a product collection; someone comparing fit may need specifications; a troubleshooting query may need a supported procedure. A blog post is one possible format, not the automatic destination for every phrase.

A fictional shortlist: packaging queries with different jobs

Imagine a new packaging supplier with a basic product collection and incomplete technical material. The five phrases below are written for this teaching exercise. They are not a live keyword export, customer case or verified search opportunity. KD, volume and CPC are all “not queried”; the proposed page choices require real research before use.

Teaching queryTask to investigateProposed next action
custom corrugated boxesBrowse products and assess supply optionsCheck whether the existing collection can show the real range and quotation inputs
corrugated packaging boxes made to orderPossibly the same task as the first phraseKeep in a candidate group; inspect results before mapping both to one URL
how to measure internal box dimensionsPrepare accurate size informationObtain an accurate measurement method and supporting visuals; evaluate a guide linked from the collection
corrugated box for chilled deliveryCheck suitability under a specific use conditionGather product and use-condition evidence before making suitability claims
custom box minimum orderDecide whether to request a quoteConfirm the supplier's actual quantity policy and applicable conditions; improve the relevant service or product page

The first two records are not automatically merged: their wording suggests a relationship that still needs inspection. The measuring question can help a buyer prepare information without producing an immediate order. The chilled-delivery query should wait for appropriate evidence if the team cannot support the intended claim. The minimum-order query requires the real supplier policy, not a number copied from another business.

Once those checks are complete, record an actual URL and owner for each task. For a proposed guide, also identify the product or support page from which readers will reach it. A sitemap entry alone does not make the page useful or integrate it into the site's navigation.

Group shared answers, not merely similar strings

Semrush's keyword-mapping guide connects query groups to existing or planned pages. That is a useful planning method, but a generated cluster still needs review. Several queries returning the same URL are a clue to shared intent, not a universal overlap percentage that forces a merge.

Keep phrases together when one clear page can genuinely complete their task. Preserve separate records where model, condition, language or next action changes the answer. A single page can include necessary supporting questions; independently useful procedures or substantially different decisions may need their own destinations. Do not prescribe “one core keyword plus five long tails” for every page.

Check the existing site before creating URLs. Improving a thin product page may be more useful than creating another near-duplicate article. Conversely, forcing every question into one large page can make a needed answer hard to find. Record the reason for the page choice, not just a color-coded cluster name.

Do not add every variant's estimated searches and label the sum unique market demand. Tools may group close variants, and query estimates are not counts of distinct people. Preserve the original definitions and conditions before attempting any aggregate.

Prioritize by readiness and value, then make a deliverable brief

A practical queue can use four states: essential page work, ready for content, needs material, and deferred. These are editorial states, not Google scores. Consider relevance, the usefulness of the answer, available evidence, execution cost and dependencies alongside the demand and competition observations.

For the fictional measurement guide, a usable brief would specify the reader's task, the dimensions and terms to explain, the verified method, the visuals still needed, the existing collection it supports, the reviewer and the publication checks. It should not promise a ranking date or ask the writer to exceed a competitor's word count by a fixed percentage.

Google's people-first content guidance asks about original information, completeness and trustworthy sourcing; it rejects a preferred word-count target. The article-length guide shows how to decide depth from the task. A narrow answer and a detailed procedure need different space because they do different work.

Keep publication and search outcomes as separate states

After publishing, verify the actual page, metadata, mobile readability, relevant links and supplied files. Then observe crawl and indexing information where available, search impressions and clicks, and qualified enquiries under their actual definitions. “Published,” “indexed,” “ranking” and “generating business” should not share one checkbox.

Revisit the list when material arrives, a page changes, important results change or enough comparable data accumulates. Preserve earlier decisions and explain revisions. There is no universal three-month deadline at which every new-site keyword has either succeeded or failed, and a changed ranking alone does not isolate the cause.

The aim is a manageable plan of useful pages supported by real information. The list is working when the team can explain what to build next, why it belongs on the site and what still needs checking—not when every row has a green difficulty score.