SEO for 2027: what changes in code and architecture when AI becomes the new results page

The article "Evolving SEO for 2027: What still needs to change," published by Adam Tanguay on Search Engine Land, starts from a simple and uncomfortable thesis: you can have a clean ranking, an up-to-date content calendar, and a keyword map that still "works" on Ahrefs or Semrush, and still lose the moment. The synthesized answer that never sends the click, or the random Reddit thread that the model trusts more than your blog, wins on the referral metric.

Tanguay is Organic Growth Lead at SpaceXAI and writes for people who work on content strategy. Here, the focus is different: what this shift in surface demands from those who write the code, define the schema, and design a site's information architecture.

What changed in the search surface

The piece lists the fronts that stopped being experiments. AI Overviews launched in the US in May 2024. Since then, Google has added AI Mode, a conversational layer within search itself, with Gemini built in, not just as a chatbot on the side. And outside Google, people have started asking ChatGPT, Grok, Claude, and Perplexity the kind of thing they used to type into the search box.

The technical consequence is that the "client" consuming your page is no longer just Googlebot rendering HTML to display a snippet. Now there's a fleet of LLM crawlers reading, compressing, and reciting your content, often without generating a visit. The job of making content machine-readable has become the job of making content machine-citable.

The numbers behind the thesis

The article draws on a Shopify analysis released in August. It's worth reproducing the data at the same scale as the source, along with the same caveat the author himself makes (Shopify didn't disclose the number of merchants in the sample, so he treats the findings as directional):

| Metric (Q2, year over year) | Result | |---|---| | Sessions from AI to stores | +197% | | Conversion of AI-driven shoppers | ~2x that of organic traffic (in high-research categories) | | Organic search growth | +12%, on a much larger base | | New customers in "taste-led" categories | ~1.3x more via AI |

Mikhail Parakhin, Shopify's CTO, said that sessions from AI "roughly tripled" while organic search kept growing on a larger base. The author's reading is straightforward: AI is a smaller slice, but one growing faster and with higher intent, while organic Google remains the main volume engine.

For developers, this kills a common temptation: putting everything into "AEO" (answer engine optimization) and abandoning classic SEO. Google is still what delivers volume. The mistake would be treating the two as separate projects.

What this demands from code and architecture

The part the article doesn't detail, but that follows directly from it, is where this outlet's reader comes in. If the goal has shifted from "ranking" to "being retrievable and citable," some engineering decisions carry more weight:

• Content in the HTML, not hidden behind JS. LLM crawlers tend to be less tolerant of client-side rendering than Googlebot already is. Content that only appears after hydration is content at risk of not being read. SSR or pre-rendering stop being a performance detail and become a condition for existing in the answer.
• Structured data as a fact anchor. Schema.org (Article, Product, FAQPage, HowTo, VideoObject) helps the machine extract the "primary fact" it will recite. Tanguay's point that models struggle to "fake" a primary source connects directly to properly marking up author, date, price, and method.
• Video with transcription and chapters. The article is explicit: search systems increasingly treat video and transcripts as source material, and YouTube frequently shows up as a cited source in AI Overviews. In practice, this means generating transcripts (transcript), chapters, and VideoObject with timestamps, so the model can retrieve the right segment. A clip made with stock voiceover "doesn't make you the primary source."

The triage framework that changes the backlog

The most actionable recommendation in the piece is to rewrite the planning questions. Instead of starting with "which keywords should we try to rank for?", he suggests taking the quarter's Semrush or Ahrefs list and splitting it into three buckets:

1. Still generates clicks (comparison, price, implementation);
2. Answer only (definitions, "what is");
3. Community-owned (Reddit, "has anyone used X?").

The guidance: stop writing shallow articles for the middle bucket, show up as a named person in the last one, and concentrate production time on the first. As he sums it up, a 2,000-word "what is" piece that the AI Overview resolves on its own won't pay off; the topical authority that still pays off is the kind that gets cited across several related prompts within the same cluster.

If the model can put together a decent summary from 10 generic posts, what it still struggles to fake is a named person who used the product and will say what broke.

-- Adam Tanguay, Search Engine Land

For content engineering teams, this is a roadmap prioritization criterion: bucket 2 pages are candidates for consolidation or for becoming short answers within larger pages, instead of new articles.

Where it's NOT worth investing

The piece is honest about the cost. Chasing every undefined ranking factor inside each model is a bottomless pit, because what triggers external citations changes fast and varies by category. Two things are flagged as a waste of time or a risk:

• Astroturfing. Filling a Quora thread with soulless posts that repeat marketing copy is worse than building an imperfect but real brand voice. Unlinked mentions in trustworthy communities are work "adjacent to SEO," not branding leftovers.
• Generic AI content for head terms. Spreading thin pages across head terms loses to real depth on a topic.

And there's a metric that developers need to bring into the conversation with the business: losing blue-link traffic shouldn't be the headline if conversion, pipeline, and CAC improve. In other words, dashboards need to stop treating organic sessions as the only KPI. Measuring AI citations, assisted traffic, and conversion by source starts to matter more than the click curve.

What remains open

The article doesn't offer a closed-form recipe for instrumenting LLM citation, and that's the practical gap for those who build: there's still no consolidated standard equivalent to Search Console for measuring "how many times ChatGPT cited me." AI visibility tools exist, but they're young. Until then, the safest bet is solid basic engineering: servable HTML, correct structured data, transcribed video, and content with a point of view that only your source can provide. As Tanguay concludes, more surfaces and more algorithms should change the planning document, not the job title. The work is still about understanding the user, now in more places.