SEO 实操手册：赢的是主题，发出去才算数

Ethan Smith started his career in user research and UX design — user testing, ethnography, UX workflows — and in 2007 he was working at a commerce company where the designs he was proud of attracted almost nobody. Ads were expensive, so out of necessity he started learning SEO in his spare time. He hired several consultants, most of whom were not good, but one who had worked at Shopping.com and eBay handed over genuine playbooks that his team then tested for about three years. Most tests failed, a few succeeded, and the iteration produced what was then called the programmatic SEO search-page strategy; he later moved into product and growth broadly, and Whirlpool eventually bought the company he was at.

Ethan had consulted on the side since 2014, starting with Thumbtack, and had always found agencies somewhat distasteful — he planned to build a product instead. MasterClass, Honey and Ticketmaster then reached out with projects, those projects went well, and inbound interest grew, so Graphite became an agency focused on SEO and content strategy, much of it editorial SEO inspired by the early MasterClass work. The differentiator he points to is that Graphite builds: getting things done is sometimes more important than the strategy, so the team implements as much of the work as possible instead of handing over a recommendation deck.

The biggest myth about SEO, in Ethan's view, is the technical audit. Companies arrive asking for a list of bugs to fix when what they actually want is to grow, and the ritual is to buy Screaming Frog for around $150, crawl the site and deliver a report of errors — which is not very useful. He reframes the request as a technical strategy combined with a general SEO strategy designed to grow the business. The related mistake is over-indexing on technical SEO while under-indexing on editorial SEO.

Technical SEO, he explains, is mostly internal link architecture plus a handful of other things and bugs — it is not how you grow. Growth comes from programmatic SEO and editorial SEO. Editorial, meaning articles and content written by an actual person, is the most underappreciated of the two, and it has gone from a minority of SEO when he started to driving more traffic than programmatic pages today.

Ethan evaluates two things before recommending SEO. First, is the addressable market large — which it is for most categories. Second, do you have authority and existing traction? Seed and Series A companies typically do not, and it is too early: Google does not want a site that exists only for SEO, it wants a credible domain, and a company starting from zero gives it no signals to believe that.

He uses back-of-envelope benchmarks for those signals. In Similarweb he looks at total traffic broken out by channel and wants roughly 1,000 visits a day that do not come from search; in Ahrefs or Semrush he wants roughly 1,000 referring domains. You can grow with less than either, but it is much harder and slower — the more authority you have, the faster and larger the growth can be.

The addressable market should be sized by product use cases and personas rather than by literal query volume. Power, a clinical trial lead-generation site, is his illustration: very few people type "I want to take a clinical trial" into Google, but the population of plausible candidates — gig economy workers, college students, people matching the key demographic — is enormous, so the content targets the persona. He also benchmarks against external reference points, putting other companies in Similarweb: a shopping site can look at Walmart and Wayfair's total traffic to see how big it could get.

He draws a distinction between product competitors and audience competitors. Robinhood competes with other brokerages, but its audience competitor is Investopedia — a site that does not let you buy stocks yet captures enormous search traffic for finance questions. Multiplying a competitor's traffic by an expected conversion rate gives an estimate of potential sales, which can then be compared with paid, social and other channels. If paid is working well or the brand is already strong, it can be rational to delay SEO; the decision is relative priority, not principle.

Ethan explains why non-search traffic is treated as an authority signal: if every visit comes from search and nothing else, you look like an "SEO barn" rather than a real brand. Direct, paid, email and social traffic all count, and the more of those signals you have — along with referring domains and shares — the more Google treats you as a known entity and the better you can rank.

The decision framework is straightforward arithmetic: expected traffic from a market multiplied by a conversion rate equals sales, which can then be compared with what the same money would return from paid or social. The second axis is impact versus scope. Having lots of user-generated content or SKUs — Glassdoor reviews, Zillow listings, a large catalog — lowers the scope of competing; having no content and no UGC raises it, though editorial SEO remains available as a path.

Ethan sorts SEO into three buckets. Programmatic SEO is pages generated automatically from a database, such as eBay product pages, Zillow address pages or Glassdoor's user-generated content, and it comes down to two main page types: category pages that are grids or lists of items, and item pages for a single product, home or video. Editorial SEO is where a writer sat down and produced an article, guide or listicle. Technical SEO is the underlying infrastructure — internal link architecture, tags, redirects and page speed — and it applies to everything.

Which bucket to favor depends on the page type that maps to the queries, but the short answer is usually editorial. Zillow and eBay should mostly generate pages programmatically, while most other companies should invest in editorial. The economic difference is that programmatic pages carry no per-page cost — Zillow does not pay for each home page it creates — whereas every editorial page does. His reference examples are MasterClass and BetterUp for editorial work he was involved in, plus NerdWallet, HubSpot and Dotdash, which he calls probably the most successful editorial SEO company.

Ethan's workflow runs in a fixed order: define the addressable market and the topics and keywords worth acquiring, decide the page type those queries require, derive the product requirements for that page, build the content strategy of subtopics and themes, and finally the infrastructure that lets Google find everything through links. A handful of page types — category, item, article or listicle, product marketing pages and tools such as calculators or Zapier's integration pages — cover about 90% of use cases.

Page type is chosen by looking at what already ranks. Search "how to poach an egg" and every result is an article containing a recipe, so that is the page type required. MasterClass had a Gordon Ramsay chapter about poaching an egg, but a chapter page with a transcript would not rank, because Google was not ranking chapter pages or transcripts — the content existed and the page type was wrong. The rigorous version of this exercise repeats the check across the thousand keywords the page should target, looking for correlations in page types rather than eyeballing a single query.

Some queries have fractured intent: "best cameras" is served by a category page and a listicle, so you need at least one of those, and a product page will not rank. The same correlation analysis reveals required components — whether the ranking pages have a map, an address, a phone number — which becomes the product requirement. Ethan insists this is inspiration rather than copying: if users want to find and contact a business, include a map, but also ask what else would genuinely help them, because usefulness is what survives algorithm changes.

Engagement is the mechanism that turns usefulness into rankings. Google watches whether people click your result versus others at the same position, whether they bounce back to Google and click a competitor — which he calls a very bad signal — and whether they stay on the page and find the answer, which Google can observe after the click through Chrome and Android. High engagement ranks you higher immediately and makes future updates less threatening.

On scale, Ethan says it is usually not millions of pages unless you are eBay or Zillow. For item pages you index everything above a quality threshold and skip thin, empty or out-of-stock pages. For category pages the old approach — one page per possible query, which gave his 2007 employer more than 10 million pages — no longer works; Google's guidance explicitly discourages many minor variations of the same query, so keywords are grouped at the topic level and only those pages are launched.

The reason is that any given page typically ranks for 200 to 2,000 different keywords, because the algorithm has enough semantic understanding to treat variations as one concept. This has three consequences. Search volume for a single keyword undercounts a topic, often by 10x or more, which leads to badly wrong prioritization; writing a thousand articles for a thousand variations wastes a large amount of money; and the page that wins must satisfy the intent behind all of those variations, including themes you may not have known existed.

Butter lettuce is the canonical example: MasterClass's page ranked for about 400 keywords spanning themes like health benefits, other lettuce varieties and recipes, and it was able to outrank Food Network — a far stronger domain in food — because incumbents left those gaps open. BetterUp's five-year plan article ranks for roughly 1,000 keywords covering templates, examples, personal versus business plans and five-year versus ten-year versus one-year plans. To decide whether themes deserve their own page, compare SERP overlap: identical results mean one page, different results mean the topics should be split.

Topic discovery is competitive archaeology. The team pulls competitors' pages and the keywords mapped to each from Ahrefs or Semrush, clusters those keywords into topics, and sorts by search or traffic volume. Then they layer on how competitive they can realistically be through topical authority, and on conversion intent — food versus comics versus photography convert very differently — to produce a prioritized list of topics ranked by expected conversions.

Ethan's tool stack starts with Google Search Console, which is free, gives real query and traffic data and includes SERP tracking for every keyword you rank for. Clearscope is his pick for content analysis and scoring, because it exposes the subtopic gaps that determine comprehensiveness. For keyword research he prefers Ahrefs over Semrush, though they are similar and cost roughly $100 to $200 a month, and Screaming Frog handles crawling other sites, auditing your own and counting internal links. Graphite also builds in-house tools for topic clustering and topical authority, which no off-the-shelf product covers well.

Expect six to nine months before SEO shows results, and treat that as a function of existing authority and traction. A company like Netflix sees a change the next day; a brand-new site takes about a year. The first roughly three months go into analysis, building page types, launching and writing about 20 articles, followed by three to six months of ranking and growth. Early on, Google tests your pages on less popular keywords and expands as engagement proves out, which is why growth compounds — and at a large, authoritative site the post-launch lag can be as short as one to four weeks.

The leading indicators come in the form of position movement on low-volume terms: ranking around 15, then 12, then 8 with impressions but few clicks, then 5 where clicks arrive. That is the proof that the same approach can work at scale — going from 10 articles ranking at position 5 to a thousand articles — and Ethan waits for it before scaling spend. He warns that most "SEO doesn't work" conclusions are false negatives caused by bad implementation, and that every claim, including Google's own statements or another practitioner's advice, should be treated as a hypothesis validated against data from your site and competitors' sites.

On hiring, Ethan is biased toward in-house even though he runs an agency. Programmatic SEO is the harder hire: it requires people who understand indexation logic, keyword clustering into browse categories, UGC indexation and scalable systems, and there are few of them. Editorial is a much larger talent pool. He also argues SEO is systematically under-resourced relative to ads — if a company will spend $100 million on paid and see comparable organic traffic, spending $50,000 on SEO does not make sense.

To find a programmatic SEO person, Ethan looks for sites nobody has heard of where SEO is an unusually high share of traffic — around 80% — and still growing, because that is a harder path than doing it at a well-known brand, and those people are less contested. The day job is not writing: it is auditing other companies' sites, pulling their URLs and keywords, writing product requirements, manually clustering keywords into browse categories, analyzing what worked, and running tests. He is blunt about sequencing: if you are starting from zero, you should be launching, not testing — testing begins around month six, and page-type or sequential tests are acceptable where domain-wide or graph-wide changes cannot be split.

On the editorial side, the DataDog story of hiring an engineer to write blog posts is directionally right, because writing genuinely useful content is the most important part. But two other steps are easy to miss: choosing which topics to write about — many interesting topics have no search demand — and making sure the article covers all of the subtopics. A strong writer may guess the outline correctly, but most do not, which is how pages end up missing the health benefits of butter lettuce entirely. Clearscope-assisted outlines close that gap.

The workflow looks like a small factory: prioritized topics, then outlines, then domain-expert writers, then publishing through a CMS such as Contentful or Webflow, then analysis feeding back into the strategy. The SEO manager owns topics, outlines, publishing and results, but rarely does the writing. Encouragingly, that manager does not need deep prior SEO experience — at MasterClass the team had none, learned Graphite's workflow, and became excellent at content strategy first and SEO second. Deep expertise matters most on the programmatic side.

Volume should be staged. Ethan typically starts at 10 to 25 articles a month and waits for the leading indicators before going deeper; once a company has proof — say 50 articles with half ranking around position five — it becomes an ROI question of cost per article against the return, and companies then push past 100 articles a month. The arithmetic of coverage drives the pace: with 20,000 topics, waiting is not an option, but neither is writing a thousand articles in month one. The smallest viable team is one SEO person plus one writer, while Graphite usually runs three to five people to move faster.

The dirty secret of SEO consulting, Ethan says, is that most recommendations never launch. When he interviews SEO people from other agencies, they describe elaborate strategies and then admit almost none of it shipped for lack of resources. He attributes this to four things: SEO usually sits outside the product org even though its changes are product changes, which is the argument of Eli Schwartz's product-led SEO; there is often no executive buy-in; the work is mysterious, and mysterious work loses prioritization to legible work; and the scope is simply large. Getting things done is frequently a bigger factor in success than the strategy itself.

Topical authority is, in his words, the most important and least understood part of SEO. PageRank began with links — more referring domains means more importance — but authority now has many inputs: non-search traffic, which Google can observe through Chrome and Android, shares and social activity, and branded search volume. Crucially it is topical rather than domain-wide: Google looks for disproportionate expertise in a theme, inferred from anchor text, from what gets shared and from queries like "MasterClass Gordon Ramsay," and then extends that authority to semantically adjacent topics. His jello-shots example makes it concrete — a viral post about jello shots not only ranked around position six, it made the whole domain rank for vodka and whiskey recipes that were never linked from it, which is how a smaller site can beat an Investopedia on a narrow theme.

Internal links are the most reliable quick win he knows. In 2007, Search Console exposed per-URL internal link counts, and correlating those with traffic showed that pages with more internal links got more traffic; when the team added body links to unlinked product pages, organic traffic to them rose by roughly 300%, and the mechanism still holds. He estimates 90 to 95% of sites have this problem because internal linking is usually facet-based, recency-based, popularity-based or driven by "related" recommendation algorithms, all of which concentrate links on a small subset — a power curve where about 5% of pages collect 95% of links. Google needs multiple paths, roughly ten links, to crawl a page often enough to credit it, so the fix is to flatten the hierarchy, shorten the distance from crawl entry points and place links in the body above the footer, since navigation and footer links are discounted. The typical opportunity is 25% or more, sometimes 100%.

On AI-generated content, Ethan is mostly negative and offers a precise reason: models generate sentences similar to sentences already written, not facts. Playing with GPT-3 to describe a soap produced the claim that it was "oxide free" simply because something else on the web was, and the text read perfectly while being false — dangerous in topics like cancer treatment or retirement planning. Where structured data exists, generated sentences can be factual and useful, as with Power extracting structure from clinical trial records. The better application of AI is everything around the writing: clustering thousands of keywords, finding subtopic themes, drafting outlines, assessing topical authority and analyzing performance, with a human domain expert writing the final piece.