Skip to content
Deepak Gupta markGEO Compass

Research · AEO + GEO · 16 min read · last updated 2026-05-27

The state of AI search, mid-2026: what has actually changed for publishers

Six grounded engines, two disciplines, one collapsing distinction between AEO and GEO, and the patterns that have stabilised enough to plan against

Most AI-search writing in mid-2026 is either evangelism or panic. Neither is useful for the team that has to ship next quarter's content. This research piece is the working snapshot: what has stabilised, what has not, and what the field reports on guptadeepak.com over the last 18 months actually tell us about how the discipline is settling.

It draws on the body of GEO and AEO writing on the blog rather than restating it. Where a section relies on a specific argument made in more depth on guptadeepak.com/posts/, that piece is linked inline.

What has stabilised

1. The six-engine landscape is the planning unit

Eighteen months ago AI search optimisation was a one-engine conversation (ChatGPT). In mid-2026 the planning unit is six grounded engines (ChatGPT Search, Perplexity, Claude, Gemini, Google AI Overviews, Bing Copilot), each with distinct grounding behaviour and citation surface. The engine profiles on GEO Compass document the per-engine specifics; the meta-point is that any program treating "AI search" as one channel is mis-segmenting its measurement.

See The future of search: AI-powered transformation for the longer arc of how this happened, and The AI revolution in search for the user-side framing.

2. AEO and GEO are converging into one practice, slowly

The bridge guide on GEO Compass treats AEO and GEO as two adjacent disciplines with a shared content foundation. The apex blog's practitioner breakdown reaches the same conclusion from a buyer-side angle.

The convergence is real but slow. As of mid-2026:

  • The optimisation tactics that overlap (schema, dating, Person/Organization graph) are now most of the work. Programs that get these right are already in the top quartile.
  • The tactics that diverge (FAQ schema and snippet shaping for AEO; llms.txt and long-form depth for GEO) are 20-30% of the effort but produce most of the marginal returns.
  • Measurement remains split. No single vendor reports both featured-snippet wins and AI-engine citation share with comparable methodology. Programs assemble measurement from 2-3 vendors.

The honest prediction: the disciplines collapse into "AI search optimisation" by late 2027. Until then, treating them as two tracks with a shared foundation produces cleaner programs than treating them as one or as fully separate.

3. Citation share is the load-bearing KPI

The measurement guide on GEO Compass argues for citation share as the headline KPI. The board-level version of that argument lives in Citation share: the metric cybersecurity CMOs should be reporting to the board in 2026, and the case that classical SEO metrics no longer predict revenue lives in Why I cancelled Semrush after 7 years.

The unifying point: citation share is imperfect, methodologies differ across vendors, but it's the only metric that ties content investment to AI-search visibility coherently. Programs that report it consistently over 18 months see what works even if the absolute numbers are noisy. Programs that don't measure cannot tell.

4. Product content beats blog content in B2B AI citations

The 768,000-citation study summarised on the blog is the single most-cited data point in our reference set: in B2B, product content earns 46-70% of AI citations while blogs earn under 6%. The implication restructures content investment: documentation, integration pages, security posture pages, methodology pages become primary GEO assets rather than supporting material.

This is the most actionable finding of the year. Most B2B teams have inverted budgets relative to what citation data suggests.

5. Gated content is invisible to AI engines

The argument is made in full in Why gated whitepapers are killing your AI visibility. Short version: MQL-driven content strategies that hide your best material behind forms produce zero AI citation share. The pipeline cost of un-gating is real; the AI-visibility cost of staying gated is larger and growing.

By mid-2026 most leading B2B SaaS companies have un-gated their highest-authority content. The laggards are visibly losing share.

What has not stabilised

1. Vendor measurement methodology

The vendor matrix tracks 18 measurement and visibility vendors as of mid-2026. The category is consolidating but methodologies still differ enough that two vendors will report different citation-share numbers for the same domain. Pick one methodology, run it consistently, trust the trend rather than the absolute number.

The category will likely consolidate to 5-7 serious players by end-2027 with a Gartner-style coverage report formalising methodology comparisons. Until then, vendor selection is a judgment call.

2. AI shopping assistants

The ecommerce vertical guide documents the rapid evolution of Perplexity Shopping, ChatGPT Shopping, Google Shopping AI, Microsoft Shopping, and Amazon Rufus. Behaviour changes month-to-month. Programs that optimise for clean Product schema fundamentals (Merchant Center, structured feeds) benefit regardless of which engine wins.

3. llms.txt formal endorsement

The llms.txt deep dive on GEO Compass treats publishing one as a no-regret move: small cost, asymmetric upside. As of mid-2026 no major AI engine has formally endorsed the spec, but publisher adoption is rising fast. The downside risk of waiting for formal endorsement is larger than the cost of shipping early.

4. The publisher–LLM economic settlement

The crisis is laid out in The AI content crisis: how LLMs are draining media revenue. Licensing deals (NYT/OpenAI, Axel Springer/OpenAI, Reuters/Meta), Cloudflare's AI bot controls, and watermarking technologies are the current settlement experiments. None is dominant yet. Publishers' AI strategy in 2027-2028 will look very different from today's.

The B2B SaaS narrative that has stabilised

The clearest year-over-year narrative is in B2B SaaS, where the operator-side argument has been made consistently across:

The synthesis: B2B SaaS buyers now use AI engines as their primary research surface for 60-80% of evaluations (industry-dependent). Content that doesn't show up in those engines doesn't reach the buyer. Classical SEO metrics undercount this because the AI traffic doesn't always click through; the buyer reads the cited sentence and forms an impression without ever visiting your domain. Programs that track only click-through traffic are blind to most of the funnel.

The B2B SaaS vertical guide on GEO Compass is the operational version of this argument.

The cybersecurity narrative

Cybersecurity is the vertical where AI-search behaviour has been most extensively documented:

The findings that generalise beyond cybersecurity:

  • Vendor-neutral writing earns disproportionate citation share. Cybersecurity buyers have heard every pitch; the vendors that win cite sources, declare conflicts, and publish honest weakness analyses.
  • CVE-driven content has unusually short half-lives. Rapid publication and aggressive update cadence matter more in cybersecurity than in adjacent B2B verticals.
  • Compliance content is a long-tail citation engine. SOC 2, ISO 27001, HIPAA, PCI DSS: each framework's specific clauses become persistent citation queries.

The cybersecurity vertical guide on GEO Compass is the operational version.

The technical foundation that everyone agrees on

The RAG architecture deep dive is the cleanest published explanation of why the structural work (schema, dating, llms.txt) actually moves the metric. Every grounded engine is a RAG system at heart: chunk, embed, retrieve, ground, cite. Content that chunks cleanly, dates explicitly, and provides citable sentences gets retrieved disproportionately.

The semantic-search and data-processing deep dive covers the publisher-side implementation patterns.

The schema-for-AEO-and-GEO guide on GEO Compass is the priority order for the structured-data layer; the citation-worthy content patterns guide is the priority order for the editorial layer.

Programmatic SEO and AI search

The programmatic-SEO body of work on the blog (the complete guide, the #1 growth hack framing, the paradox piece) covers a tactic that is both more powerful and more risky in the AI-search era.

More powerful because grounded engines reward structured, comparable content at scale (vendor matrices, comparison pages, category landing pages). More risky because AI engines detect low-effort scale faster than classical search did. The pattern that works in mid-2026: programmatic scale with editorial quality at the page level (structured data, dated, with at least some genuinely original sentence per page).

What this means for next quarter

If you run a B2B content program in mid-2026, the priority sequence that the field has converged on:

  1. Product content depth before blog volume. Documentation, methodology, integration pages, security posture, public roadmap. The pages that are usually treated as "supporting material" earn the citation share.
  2. Un-gate the highest-authority content. Pipeline cost is real; AI-visibility cost is larger.
  3. Schema and llms.txt foundation. Three weeks of work for a typical site. The marginal return per week of effort is highest here.
  4. Citation-share measurement on a defined query set. 100-500 queries, run consistently for 18 months. Trend reveals what's working; absolute numbers are noisy.
  5. Editorial discipline on citable sentences. Every paragraph carries at least one quotable claim. This is editorial work, not SEO work.
  6. Vendor selection for measurement. Pick one methodology; trust the trend.

The teams shipping this sequence in 2026 are 18-24 months ahead of teams that wait for AI search behaviour to "settle." It has settled enough.

Open questions for late 2026 and 2027

  • Does Google AI Overviews' citation behaviour stabilise enough for confident optimisation, or remain a moving target?
  • Do AI shopping assistants reach a tipping point where SKU-level GEO becomes its own discipline?
  • Does the publisher/LLM economic settlement converge on licensing, on Cloudflare-style bot controls, or on a watermarking/attribution standard?
  • Does llms.txt get formal endorsement from one of the major engines, accelerating adoption further?
  • Does measurement methodology consolidate enough for cross-vendor citation share comparisons to be meaningful?

The next research update on GEO Compass will revisit these questions on a quarterly cadence. The blog will continue to ship field reports as they're written.

Related guides

Further reading on guptadeepak.com

← All guides