Search did not die when generative answers arrived. It became less forgiving of content that says the same thing as everybody else.
That distinction matters. A great deal of “GEO” advice in 2026 is ordinary SEO wearing a more expensive jacket: add FAQs, mention entities, publish more pages, hope an AI quotes you. Google’s own guidance is calmer and more useful. Make the site crawlable. Publish material that is unique, first-hand and genuinely useful. Give both people and machines enough context to understand who created it, what it covers and why it should be trusted.
The opportunity is real. Google says AI Mode has passed one billion monthly users, that its queries have more than doubled every quarter since launch, and that the average AI Mode query is three times longer than a traditional search query. People are not just typing a shorter keyword into a new box. They are describing a situation, constraints and a decision they need to make. Read Google’s 2026 AI Mode usage findings and the official generative AI search optimization guide.
For a startup, this changes the content job. Ranking for a phrase is useful. Becoming the clearest evidence source for a decision is better.
The Part That Did Not Change
Before inventing a new AI-search department, get the foundations right.
Your important pages still need stable URLs, descriptive titles, crawlable links, useful text, fast rendering, sensible canonicals and a sitemap. JavaScript can enhance discovery, but core articles should exist in the initial HTML. An AI system cannot confidently retrieve a page that a crawler cannot reliably reach.
Internal linking still matters because it explains relationships. A pricing article should connect to unit economics, packaging and customer research. A product-feedback article should connect to onboarding, retention and prioritisation. These are not decorative “related posts” links. Together, they describe the knowledge graph of the site.
Authority still has to be earned. Structured data can explain that an article was written by Evan D’Souza, but it cannot prove that Evan has operated startups. The proof comes from consistent authorship, an accessible professional record, concrete examples, original frameworks and claims that survive inspection.
Google says the same thing more directly: unique, non-commodity content is more likely to influence long-term visibility than content that merely restates information already available. It explicitly contrasts first-hand experience with generic summaries that a model could produce. That is the line every article on this site should now clear.
The Query-to-Evidence Model
Longer AI searches usually contain four layers:
- Situation: “We are a 12-person B2B SaaS company.”
- Constraint: “We cannot hire a full growth team this quarter.”
- Decision: “Should we invest in SEO, partnerships or outbound?”
- Quality bar: “Give me a plan with metrics and trade-offs.”
A page built around the keyword “B2B SaaS growth” is too broad to resolve that request. A useful page has to help the reader make the decision. It needs a diagnostic, a comparison, a worked example and an implementation path.
I use a simple standard for every serious article:
A page is decision-complete when a qualified reader can choose a next action without opening five generic tabs.
That does not mean cramming every possible answer into one page. It means satisfying the real decision behind the query and linking clearly to the next layer of depth.
The Seven-Part AI Search Audit
1. Crawlability before cleverness
Render the article content as HTML. Use real anchor links. Keep canonical URLs stable. Include articles and topic hubs in the sitemap. Do not hide the archive behind a search box or a button a crawler has to click.
On this site, the article library now exposes five intentional, crawlable collections instead of treating dozens of ad hoc tags as independent destinations. Filters are useful for people, but unrestricted faceted URLs can generate duplicate combinations and waste crawl attention. Google’s faceted-navigation guidance recommends controlling that URL space.
2. One real decision per article
“Everything about AI” is not a topic. “How should an early-stage team divide a support workflow between humans and an agent?” is.
Write the primary decision at the top of the content brief. Then list:
- who is making it;
- what evidence they need;
- what can go wrong;
- what action should be possible after reading.
If the article cannot answer those four questions, it is probably a keyword container rather than a useful resource.
3. Build evidence blocks
An evidence block is a claim that can be lifted, understood and checked without losing its context. It normally contains:
- the claim;
- the source or first-hand observation;
- the conditions under which it holds;
- the implication for the reader.
For example: “Average AI Mode queries are three times longer than traditional queries” is a data point. The implication is that content should resolve richer situations and constraints, not produce hundreds of pages for tiny keyword variations.
Original evidence is even stronger. A screenshot of an operating dashboard, a real decision template, an anonymised workflow or a postmortem from something that broke creates information that did not exist before the article.
4. Make the author entity consistent
Use one professional name, one canonical author page and consistent same-as links to credible profiles. Connect articles to the author’s actual areas of knowledge. Avoid changing the biography on every page to chase the topic of the day.
For Evan, the durable entity is not “AI expert” or “growth hacker”. It is startup operator, consultant and builder working across operations, growth and product. AI is part of how the work is executed, not a costume applied to the brand.
5. Add media that teaches
Google’s AI-search guide specifically calls out high-quality images and video as additional opportunities for visibility. The operative phrase is high-quality. A decorative robot holding a magnifying glass teaches nothing.
Useful media includes:
- a labelled framework;
- a before-and-after workflow;
- a short explanation of a decision model;
- a real interface or dashboard with sensitive data removed;
- a downloadable template.
Each asset needs a descriptive filename, useful alt text and nearby text that explains what the reader is seeing. The image should still make sense when the social crop changes.
6. Organise the library around problems
Tags accumulate. Taxonomies are designed.
This article library previously surfaced roughly 60 distinct tags at once. That made the archive look extensive but forced the reader to understand the site’s internal vocabulary before finding an answer. The new structure uses five durable operating lenses: Operating Systems, Growth & Brand, Product & Customer, AI & Building, and Founder Field Notes.
Secondary tags still help with filtering. They no longer pretend to be the information architecture.
7. Measure assisted discovery, not only blue-link clicks
Search Console includes traffic from Google’s AI search experiences in the overall web reporting described in Google’s guide. Track impressions, clicks, landing pages and query families, but do not stop there.
For a consulting site, the useful sequence is:
- organic or referral landing;
- engaged article read;
- movement to a related article, proof page or consultation page;
- qualified enquiry.
A page can earn fewer clicks and still create better commercial outcomes if the visitor arrives with stronger context. Measure the content path, not just the first hit.
A Worked Content Brief
Suppose a founder wants to rank for “AI customer support for SaaS”. The commodity article is a list of tools. The decision-complete version is different.
Decision: Which support work can an AI agent own without damaging customer trust?
Reader: Founder or Head of Support at a B2B SaaS company with 500 to 5,000 monthly conversations.
Evidence required: query mix, resolution quality, escalation frequency, knowledge coverage, risk by ticket type and cost per resolved issue.
Original artifact: a workflow matrix dividing tickets into auto-resolve, draft-for-review, human-owned and prohibited.
Worked example: one billing dispute, one account-access request and one product question moving through the matrix.
Conversion: download the evaluation checklist or discuss the operating model.
That page has a chance of earning attention because it contains a decision system. A list called “10 Best AI Support Tools” can be reproduced before lunch.
The 30-Day Publishing System
Do not publish 30 articles in 30 days. Build one small body of work that compounds.
Week 1: Map demand. Pull real questions from Search Console, sales calls, support conversations, communities and founder discussions. Group them by decision, not word similarity.
If those signals are scattered across support, sales and product tools, build a voice-of-customer operating loop before commissioning another content calendar. The language customers use is usually a better demand map than a brainstormed keyword list.
Week 2: Build one flagship resource. Include the framework, example, limitations, source links, FAQ and one useful visual.
Week 3: Publish two supporting field notes. Answer narrower questions the flagship introduces. Link both ways.
Week 4: Distribute and learn. Turn the strongest insight into a LinkedIn post, short video and email. Record which framing earns qualified responses. Update the flagship only when the substance improves.
This is slower than mass generation and much faster than repairing a domain full of forgettable pages.
What I Would Not Do
- I would not create a separate page for every AI-generated query variation.
- I would not add schema that does not describe visible content.
- I would not buy mentions from irrelevant listicles to inflate a domain metric.
- I would not quote statistics without the original source and conditions.
- I would not call a rewritten vendor blog “research”.
- I would not measure success using Domain Authority alone. DA is a third-party comparative metric, not a Google ranking score.
Strong links follow strong work. Original tools, templates, research, informed disagreement and first-hand operating lessons are far more linkable than generic advice. The best authority strategy is to create something another credible writer genuinely needs to reference.
Practical Checklist
Before publishing, verify:
- The article resolves one identifiable decision.
- The headline and description describe that decision plainly.
- Every important claim has a source, example or first-hand basis.
- The author and organisation details are consistent.
- The page is server-rendered, canonical and internally linked.
- The visual teaches or sharpens the argument.
- The article links to one deeper and one adjacent resource.
- The FAQ answers real objections, not keyword variants.
- The next action is useful even if the reader never becomes a client.
AI search did not make quality obsolete. It raised the cost of publishing something with no information gain.
FAQ
Is GEO different from traditional SEO?
The useful parts overlap heavily. Crawlability, clear structure, original evidence, strong authorship and useful media matter in both. Generative search increases the value of decision-complete, first-hand content, but it does not create a secret technical layer that replaces SEO fundamentals.
Do startups need special schema for AI search?
No special “AI schema” is required. Use supported structured data that accurately describes visible content, such as Article, BreadcrumbList, Person and CollectionPage. Schema clarifies meaning; it does not manufacture authority.
Should AI write startup blog posts?
AI can help research, organise, challenge and edit. It should not invent experience, statistics or customer stories. Google’s generative AI content guidance focuses on whether content is helpful and whether scaled generation is being used to manipulate rankings.
How many topic categories should a startup blog have?
Use the smallest set that remains stable as the library grows. For a focused expert site, four to seven primary categories is usually easier to understand than dozens of tags. Secondary tags can still support filtering without becoming indexable landing pages.
Does Domain Authority improve rankings?
Domain Authority is a third-party estimate used for comparison. Google does not use that named metric. Improve the underlying signals instead: publish reference-worthy work, earn relevant links, maintain technical quality, demonstrate real expertise and build a coherent body of content.
