AI 摘要吃掉顶部漏斗之后，SEO 该当成产品来做

Schwartz traces the shift to ChatGPT's arrival at the end of 2022, which spawned the idea that nobody would need Google or SEO any more. He does not believe that, but he says Google had to respond because one of its main stakeholders is Wall Street. If investors decided Google was a has-been, its stock, its hiring and its ability to invest would all suffer.

Google's first answer, the Bard demo, went badly. It then built the Search Generative Experience, essentially ChatGPT inside a search result, and held it in beta for about a year because of three worries Schwartz lists. Ads fund most of its revenue and sit awkwardly beside an AI answer, Google becomes the publisher liable for harmful output, and there is a plagiarism question.

Renamed AI Overviews, the feature launched at Google I/O in May and promptly produced embarrassing screenshots, such as advice to put glue on pizza. Google rolled it back but has since relaunched it on far more results, including for logged-out and incognito users and in the UK. Schwartz says he now sees it on most of the searches he runs.

Schwartz says the problem is not that your pages become harder to find but that they become less relevant. Classic SEO fixated on winning a big top-of-funnel keyword with a long piece of content, such as the best beach vacation near the US. Now the AI answer suggests a destination like Cancun, and the user's deeper searching starts from there, so ranking first at the bottom of the page matters far less.

He expected an apocalypse and admits it has not shown up in traffic yet. He points out that AI Overviews were previously limited to logged-in users and had been scaled back, and that algorithm updates muddy any measurement. Impact will depend on where the user sits in the funnel: a WebMD-style site with generic health content is exposed, while an e-commerce site may not be.

He distinguishes structured from unstructured data. Lenny's old Hanukkah-date site was killed years ago by Google answering structured queries directly, but now Google and other LLMs can turn unstructured content into statistics and summaries. Review sites like G2 and Capterra, which list the top tools for a category, are in the line of fire, whereas searches like a product's price or capacity still send users to a site.

Schwartz argues that SEO has been wrongly handed to marketers after the product is built. They are asked to make an existing product fit what search results look like, and nobody asks what the searching user actually wants. He says product managers should think about how to position the product for people who find it through their own self-discovery journey, not through social, ads or a trade show.

His earliest example is Zapier. People did not know to search for Zapier, but they knew Gmail does not connect to Salesforce, so the team built pages for every app pairing, creating a flywheel of what else a given tool can connect to. Schwartz notes that it took a couple of years before results showed, because nobody was searching for the product itself.

With Tinder, Udi Milo, then chief of growth, pushed to do SEO even though all inbound came from the brand name. The team realised Tinder solves loneliness, so it built local pages for online dating in cities worldwide, offering date ideas and Tinder as the way out of being lonely in a new place. Schwartz says AI Overviews change nothing here, because a lonely user in Dubai still needs the mid-funnel solution.

Schwartz's step one is the one he says almost everyone misses: be the user. He recalls that SurveyMonkey staff were given accounts at onboarding but rarely ran a survey, which left zero customer empathy. Step two is deciding what asset to create, often programmatic rather than editorial, and step three is building it with design, engineering, user research and a product manager.

He then argues most SaaS tools should not do SEO. At Mixpanel, following the standard recipe of keywords, content and links did not convert, because the product is expensive, integrates across a company, and cannot be bought on a credit card after one search. The problem a SaaS solves is often unknown to the buyer, which is a brand challenge rather than a search one.

He adds that Google Cloud, which has only two real competitors, should not be investing in SEO at a grand scale, because purchase decisions involve committees. Lenny summarises that such products need a sales motion and not a product-led SEO motion, and Schwartz agrees.

Schwartz wants to retire the myth that SEO is free. It costs time, resources and direct spend: an agency is likely to cost upwards of $10,000 a month, plus a CMS, engineers, design and content. At SurveyMonkey, organic traffic generated a couple of hundred million dollars a year, but through a freemium product where search led straight to sign-ups and templates.

His test is whether a given investment, say a million dollars a year, would make that money back soon, compared with the same sum spent on brand ads, influencers or paid marketing. He says SEO pays back when it is the right fit, but startups need payback quickly. He tells of a gardening SaaS weighing a $15,000-a-month content proposal when its customers were found at $10,000 gardening-show booths.

Lenny adds that there are four core growth engines, SEO, paid, virality and sales, and that large companies eventually do them all. For earlier-stage companies with limited resources, Schwartz thinks some businesses should never do it at all, and even suggests most restaurants may not need a website because their customers order through Google Maps, DoorDash or Uber Eats.

How long SEO takes and what a good conversion rate looks like both depend on the business, Schwartz says. Large known brands such as Tinder see revenue as soon as a new page type goes live, while smaller companies may wait months for Google to notice them. The key is to choose the conversion that matters, such as MQLs for SaaS or page views for a media business, instead of top-of-funnel rankings.

He says traffic that does not convert has little value. He once advised an HR-space marketplace to delete content aimed at the side of the market it did not monetize, and says the belief that traffic itself earns Google's trust is mostly a misconception. Paid marketers speak in terms of efficiency, he notes, and SEO needs the same discipline; in investor decks, traffic alone is a total waste of time unless you are a media company.

Treated as a product, SEO can be run on milestones: a month of ideation, a month to write a PRD, then shipping and an expected number of pages indexed. As an example of fast results, he describes Quora, which roughly quadrupled traffic within three months after showing answers to logged-out visitors and giving crawlers a categorised HTML sitemap to navigate.

Schwartz says Canva's templates work because Canva monetizes what people build from them, and SurveyMonkey's survey templates worked because they fed its survey product. Copying someone else's programmatic strategy, or doing programmatic for its own sake, does nothing. Anything you build for SEO has to be a piece of the buyer journey, or the user journey simply stops.

On AI content, he says AI is a tool and not a solution. Using it to churn out the same worthless filler people used to buy on Upwork is just a cheaper way to publish bad content. Using it for something like product descriptions on an e-commerce site is fine, since shoppers on a Macy's page are looking for the shoes, not essays about them.

He notes that Google's documentation says AI itself is not the problem; helpfulness is. He argues users dislike content that is announced as AI-written, citing CNET, but sees little issue with automated summaries reviewed by an editor, like earnings-report recaps. Google is also struggling with a flood of new content that is expensive to crawl, which helps explain its recent clean-up updates.

Schwartz sees AI Overviews as a branding exercise. The links inside them are partly liability protection for Google, yet in his informal LinkedIn survey of a couple of hundred people about half said they click them. If your link appears, you probably already rank below it, and if your brand is named, as in a list of top CRM tools, your brand efforts are working.

He calls link building the biggest SEO myth when it means buying guest posts on low-authority sites nobody reads. In his view the right way is to build a product and brand people mention, since Google can now read a mention without needing an HTML link. Whatever you do for PR you should do for SEO, and the two should be one effort.

Schwartz defines programmatic SEO as combining data sources into pages instead of writing long pieces by hand. He cites TripAdvisor, which merged hotel and city data with user-generated content into one page per property, and Zillow, which combined government data, comparable sales, realtor photos and neighborhood and school data. For Zillow, search is the only way anyone finds a given home's page, since nobody pays for traffic to a single house.

Editorial would mean writing a long article for every hotel or half a billion properties, which would be cost-prohibitive and also wrong for what the user wants. Whether to go editorial or programmatic comes down to what the user is looking for, such as a rating or the value of a home. He says this is also why G2, Capterra and Forbes-style listicles are being disrupted.

Programmatic is the right choice when there is scale and a user use case. He gives the older example of a page for every zip code, which users do not search for, and today's mistake of making a page for each of 10,000 software tools combined with a city. SurveyMonkey had a fat head of templates but no long tail, so it could make a hundred templates, not ten thousand.

Schwartz read the 286-page ruling in the case against Google and says the writer understands SEO and digital marketing better than most marketers. The verdict says Google has 98% of mobile searches, and rejects arguments that social media or Bing are real comparisons. For him the key takeaway is how much default deals, such as with Apple and Chrome, and force of habit drive users to Google.

He cites Neeva, a search engine from a former senior Google person, which could not succeed despite a better experience because it lacked distribution. He also points out that Bing reportedly offered itself to Apple for free and was refused, and that DuckDuckGo, which he measured at about 1% years ago, is at about 2% in the verdict despite heavy spending. He concludes ChatGPT, Perplexity or Claude are unlikely to dethrone Google on quality alone.

Lenny adds a point from Ben Thompson's analysis that the Apple deal is a revenue share any rival could offer, though nobody can match the scale. Schwartz does not think breaking up the defaults would change user behaviour, since people would still go to Google, and says the only route is to prop up a competitor. Google fights hard on AI Overviews both for Wall Street and because losing small percentages to ChatGPT costs real money.

Schwartz's main message is that SEO is not dying, because people will always want to choose rather than receive one answer, which he says is why voice assistants never took off. Search volume will shrink at the top of the funnel, but hotels still need bookings and shoe sellers still sell shoes. Information publishers like WebMD are different, since Google now gives their curated content away without ad clicks.

Among the myths he lists are that every funded startup needs SEO, that link building is the secret sauce, and that Google is a black box. Google's best-practice guide says to build a site Google can understand, link between pages and write helpful content, and he says you can get from zero to ten by reaching step eight with those basics. Technical SEO and page speed are usually overhyped for small sites, he says, because traffic losses after algorithm updates come from unhelpful content, not from site audits.

On TikTok and Instagram, he notes the Google verdict figure that 63% of Gen Z use TikTok to search, but says top-of-funnel discovery may move there while booking flights and hotels will still happen on Google. Mid-funnel is where intent exists, and traffic is lower but the metric is closer to business results. Lenny notes that at Airbnb PMs without SEO backgrounds had a major impact, which Schwartz says shows how far basic SEO can go.

Working at Faire on a launch in a new country, Schwartz developed a top-down approach to SEO forecasting. The usual bottoms-up method takes keyword volumes, an estimated ranking and click-through rate and a gross-up, and usually gives a number too small to justify funding. He proposes starting from population, narrowing to the people who would buy the product, then those who buy online, and applying a market-penetration target, average order value and purchase frequency.

The top-down model may be inexact, he says, but every assumption can be adjusted later. Keyword tools, by contrast, can anchor a forecast on a wrong first number; for WordPress, the tool figures were completely off compared with Google Search Console. He explains that tools like Semrush, Ahrefs and Similarweb estimate from proprietary algorithms and can be off by ten times in either direction.

He advises using such tools only for normalisation, for instance whether people spell a word with or without a space, and not as a source of truth. He recalls a public company whose board member, reading an external tool, claimed they were being crushed in COVID, while Google Search Console showed traffic had quadrupled. He closes by predicting growing demand for SEO expertise and advising aspiring growth advisors to sharpen communication, sales and proposals while keeping their day job.