--- title: "The State of SEO, AEO & GEO" description: "How AI search decides which content to cite, and what to do about it." url: https://research.momenticmarketing.com/state-of-seo-aeo-geo-unhyped-2026 author: "Tyler Einberger" publisher: Momentic Research published: "May 2026" --- # The State of SEO, AEO & GEO *How AI search decides which content to cite, and what to do about it.* ## The State of SEO, AEO & GEO *Momentic Research · May 2026* How AI search decides which content to cite, and what to do about it. ## AI pulls information about your brand from two sources that update at different rates. *How a model gets information about you* Training data is locked in until the model is retrained, so showing up there takes years. Live search returns whatever the index holds at question time, so fresh work surfaces in weeks. That second source is where a quarter of focused work can pay off. ## Same conclusion, almost entirely different sources. *Why two indexes matter* Across 730,000 response pairs, Google's AI Overviews and AI Mode reached the same conclusion 86% of the time while citing almost none of the same pages. A brand that performs well on one platform may not perform on the other. - **Semantic similarity: 86%** — the two answers reach the same conclusion - **URL overlap: 13.7%** — of the citations are the same *Source: Ahrefs AI Search Benchmark, Q4 2025 – Q1 2026.* ## Training data refreshes between model releases. The live index refreshes on every recrawl. *Refresh rates compared* AI Overview content rotates roughly every two days. The meaning stays about 95% the same. ## A question becomes an answer in twelve steps. *The answer chain* The same retrieved candidates feed two outputs. After step six the system either composes a page of blue links or selects passages to feed an answer model. Most optimization work targets the first path, while the second path is what AI search returns. ## From query to either a results page or a written answer. *Step through the chain · pick a path at the fork* **Read the query** The system tokenizes the words, converts them into a vector representation, and infers the underlying intent. **Rewrite the query** Query expansion rewrites the question into around a dozen variants the original wording never contains. AI Overviews appear on 9.5% of single-word queries, 46.4% of 7+ word queries. **Pull candidates** The system retrieves around 20 results per variant. With a dozen variants, the candidate pool reaches roughly 240 to 600 pages. **Clean the pool** Duplicates are removed and near-duplicates are flagged, leaving a smaller pool of distinct candidates. **Rank what is left** A fast, inexpensive ranker scores the pool first. Top candidates pass through to slower, more accurate rankers in stages. **Filter for eligibility** Safety, freshness, and policy checks apply. Around 38% of survivors also rank in the SERP top 10; 31% don't rank in the top 100 at all. The chain forks here. **Traditional search** - *Compose the page* — Survivors are arranged into the layout that ends up on the results page: blue links, snippets, featured modules. - *Write the snippets* — Titles and descriptions are generated, the previews shown before any click. - *Send the links* — A page of links and modules is delivered. The user may click one, or none. **AI search** - *Pick the evidence* — From the same pool, the system selects the subset of pages that together support an answer. - *Pull the passages* — The quotable lines and their attribution are extracted. This package becomes the grounding context. - *Write the answer* — A generator model produces the answer, constrained to the selected passages. It is prevented from making claims outside that grounding context. - *Cross-check it* — A verifier model compares each claim against the cited evidence, flagging mismatches for correction. - *Score the draft* — Additional models score the draft, marking which claims they support and which they would challenge. - *Send with citations* — The answer is delivered with citations attached. Most users skip the citation list. *Source: Composite: Google patents + Ahrefs benchmark + Momentic Studio.* ## Before retrieval, the query gets rewritten into around a dozen variants. *Query expansion* Before retrieval starts, the question is rewritten into around a dozen variants the original wording never contains. The longer the query, the more aggressively it fans out: AI Overviews appear on 9.5% of single-word queries and 46.4% of queries with seven words or more. ## Twelve rewrites, hundreds of candidates, around five citations. *Type a query · the system rewrites it into around a dozen variants* ## A page clears four checkpoints before quality scoring begins. *Eligibility filters* By the time anyone evaluates which page is better, the system has filtered most pages out. Missing any one of the four is enough to keep a page out of the answer, no matter how good the rest of the work is. ## Reachable: a bot can fetch the URL. *Checkpoint 1 of 4* The common failure usually isn't robots.txt. Cloudflare bot management, a WAF rule, or a hosting firewall may return 403 to AI user agents while most audit tools still report the page as crawlable. Audit the 499 rate in your access logs first. ## Three layers between the bot and the page. *Why the bot gets a 403* robots.txt may allow the bot while the firewall and the CDN still block it. ## Readable: the words exist as HTML rather than as a script the bot has to execute. *Checkpoint 2 of 4* A single-page app that renders every paragraph client-side returns a mostly empty shell to crawlers. The page looks fine in a browser while the bot reads almost none of it, so it tends not to enter the candidate pool. ## Quotable: a passage stands alone if lifted out. *Checkpoint 3 of 4* When every paragraph builds on the one before it, no individual passage works as evidence on its own. Retrieval still pulls the page, but the evidence-selection step can't extract a chunk that makes sense without its neighbors. ## A sectioned page produces multiple citable passages. *Wall of prose vs sectioned page* A wall of prose yields zero citable chunks, while a sectioned page yields one per section. ## Consistent: the claims line up with other sources. *Checkpoint 4 of 4* The verifier step cross-checks each claim against other retrieved sources. A page that contradicts the knowledge graph, uses a number few others use, or tells a recently rewritten brand story tends to get dropped here. This is what entity work and schema target. ## A fake brand, three planted falsehoods, six AI assistants. *When brand info is wrong online* - Worst case: Gemini and Perplexity repeated the planted misinformation in 37–39% of brand answers. - Mid case: Copilot, Grok, and AI Mode mixed manipulated content with official sources, but never reliably cited the brand's own FAQ. - Best case: ChatGPT-5 and ChatGPT Thinking used the brand's FAQ as canonical. Claude didn't reproduce the falsehoods, but didn't surface the official website either. ## Freshness gets the page into the context window. *Force 1 of 3 · freshness* Pages that haven't been updated tend to fall out at the eligibility step, before any quality scoring. The grounding window is around 13 weeks; the broader recency signal lasts about a year. **89.7%** — of ChatGPT's top-cited pages were updated in the last year - updated last year: 6.4 % - two years out: 1.4 % *Source: Ahrefs benchmark · Freshness Cliff report.* ## Sources rotate constantly while the answer barely changes. *AI Overview stability* Across 43,000 keywords, AI Overview content changed in roughly 70% of consecutive observations and 45.5% of sources were new each time. Cosine similarity between consecutive overviews averaged 0.95. A brand that drops out of one snapshot tends to reappear in the next, without anything having changed on its page. *Source: Ahrefs benchmark.* ## Citation odds hold steady, then drop sharply after week 13. *13-week eligibility window* After about week 13, most grounding pipelines flag the page as stale. ## Specificity is what the verifier step keeps. *Force 2 of 3 · specificity* Claims backed by named brands, numbers, dates, and people are far more likely to survive the verifier. Generic phrasing fails the check and gets dropped, even when the page reaches the verifier intact. **59%** — of AI Overviews contain no brands or entities at all - entities per AIO: 1.3 - entities per AI Mode: 3.3 *Source: Ahrefs benchmark.* ## The verifier removes generic claims and keeps the specific ones. *What the verifier keeps* The verifier keeps passages anchored to named brands, numbers, dates, and people. ## Structure decides whether the page can be cited in pieces. *Force 3 of 3 · structure* Word count has almost no correlation with citation. The average cited page is 1,282 words, and 53.4% of cited pages are under 1,000. What does correlate is whether individual passages can be cited on their own. A short, well-sectioned article tends to beat a long unbroken essay on the same ground. ## The average cited page is 1,282 words, and most are shorter. *Word count distribution* Of 174,000 pages cited in AI Overviews, 53.4% are under 1,000 words and only 16% are over 2,000. Correlation between word count and citation is near zero. | | | |---|---:| | Under 350 words | 16.6% | | 350 – 1,000 words | 36.8% | | 1,000 – 2,000 words | 30.6% | | Over 2,000 words | 16% | *Source: Ahrefs benchmark, 174,000 cited pages.* ## Inside any page, only a handful of passages get cited. *What gets cited · pick before reveal* ## Your brand, to AI, is the average of every piece of content about you. *How AI represents your brand* The system processes a cloud of content about you, each piece converted into a vector. Your representation sits at the center. Distinctive content shifts that center somewhere harder to mistake for a competitor. Templated content drags it back into the crowd. ## Scale content the wrong way and your center drifts off your category. *Centroid drift · drag the slider* *Source: Momentic · Influenceable Web report.* ## Two years ago, the bar was getting scraped at all. *Era 1 of 3 · how the bar has moved* The constraint was being in the training data: crawlable, indexed, and live before the training run finished. Public reranker tests still measure a roughly 57-point gap between real brands and made-up ones. Today, 28.3% of ChatGPT's top-cited pages have zero organic visibility on Google. ## Last year, the bar was being close in vector space to the query. *Era 2 of 3 · how the bar has moved* Embedding similarity took over from raw keyword matching. Pages with no shared keywords could still match a query; pages containing every keyword could still miss it. Most current optimization advice still works at this layer. ## This year, the bar is being reachable across the knowledge graph. *Era 3 of 3 · where the bar is now* The system traverses relationships between entities and selects the path with the most support. About 38% of AI Overview citations also rank in Google's top 10; 31% don't rank in the top 100 at all. The system relies on the entity graph more than on the rankings. ## Seven steps from query to cited vertex. *Era 3 · explore the traversal* > Marketing has long been about moving someone from seen, to believed, to chosen. > Most AI-visibility dashboards focus only on the first of the three. Being cited is where the work begins. > — Wil Reynolds ## Three layers between visibility and revenue. *Shots on goal* Citations are a shot on goal, not a goal scored. Track branded search and pipeline behind them. ## Around 90% of your citations come from outside the website you own. *The wider web carries most of it* Pages that do rank have a median Domain Rating of 90, among the most authoritative on the web. The website is still the foundation; most of the recommendation gets assembled across pages you don't own. **90%** — of brand citations come from off-site pages - cited pages with zero organic visibility: 28.3 % - cited pages a brand can influence: 32.3 % *Source: Ahrefs benchmark · Influenceable Web report.* ## For “best X” queries, 43.8% of sources are third-party listicles. *Where top-of-funnel recommendations come from* Across 26,000 source URLs, the most cited page type was a “Best X” blog list: software 51.4%, agency 40.6%, products 39.5%. Brands recommended also ranked higher on those lists. *Source: Ahrefs benchmark, top-of-funnel queries.* ## Get into the roundups your category already uses. *Where AI looks for top-of-funnel recommendations* - Listicles drive third-party citations. Position 1–3 converts to recommendations roughly 4× more often than position 8–10. - Landing and product pages drive first-party citations, when the brand itself is the topic. - Homepage and category pages get cited rarely, around 5% of sources, since they tend to be too generic to function as evidence. ## Most of this year's loudest advice was sold by people who profit from it. *How to weigh advice* - Scaling content with AI: pushed by tool vendors and volume retainers. Original work that few competitors publish is what moves the centroid. - Citation share as a KPI: pushed by the dashboards that sell the count. Link rates run from 10.7% on AIO to 51.6% on Perplexity, so report lift and correlation, not the raw count. - Build your own tools: a speaker promised five open-source tools and shipped none. Stick with whatever already runs reliably. ## When an AI Overview appears, position 1 loses 58% of its clicks. *What AI Overviews do to organic traffic* Across 300,000 keywords, position 1's CTR dropped from a forecast 0.037 to an actual 0.016. The hit runs across the whole page: position 5 lost 32.6%, position 10 still lost 19.4%. *Source: Ahrefs benchmark, 300,000 keywords.* ## Six moves, in order of return. *Where to start* - Open the live channel first. Audit robots.txt, firewall, CDN, and the 499 rate. The rest of this list won't pay off while a bot can't fetch the page. - Make every passage liftable. A heading and a self-contained answer per section, with named numbers and sources. - Refresh on a real schedule. News in hours, shopping in days, evergreen guides around 13 weeks. - Earn placement on the listicles your category uses. Around 44% of top-of-funnel sources are best-of lists. - Show up on YouTube. Mentions correlate with AI visibility at ~0.74, higher than any other factor measured. - Translate every metric to a dollar. Rate × volume × value × win rate equals pipeline. ## The work AI search keeps tends to be the work an honest reader would respect. *Where this leaves you* Clear writing on topics the team knows. A structure that lets claims be cited on their own. Real presence on the websites buyers trust. A refresh cadence honest about what's current. The technology keeps getting more sophisticated, while the bar for citation has barely moved in two years.