# AEO vs SEO: what the studies actually measure

Canonical: https://mudpie.ai/blog/aeo-vs-seo-what-the-studies-actually-measure/
Breadcrumb: [Home](https://mudpie.ai/) / [Field notes](https://mudpie.ai/blog/) / [AEO vs SEO: what the studies actually measure](https://mudpie.ai/blog/aeo-vs-seo-what-the-studies-actually-measure/)
Author: Ali Abouelatta (https://mudpie.ai/authors/ali-abouelatta/)
Published: 2026-09-19
Updated: 2026-09-19
Research type: research-article
Method: Comparative reading of official guidance, observed user behavior, a large agency cohort, and the original academic GEO benchmark. Every load-bearing number is paired with its sample, date, unit, and limit.

SEO still decides whether Google can retrieve, index, and rank a page. AEO adds a second measurement problem: what an answer engine says, cites, and sends back to a user.

The vocabulary change hides a different measurement problem. The studies measure different units, under different conditions, with different levels of causal confidence. A team that uses an AI-visibility score as a traffic forecast will make a bad decision even if the score itself is calculated correctly.

## The four studies to read first

| Study | What it actually measured | Decision it supports | Limit |
|---|---|---|---|
| [Google Search Central](https://developers.google.com/search/docs/appearance/ai-features) | Eligibility and official Search behavior. Google says existing SEO fundamentals remain relevant, there are no additional technical requirements, and no special schema is needed for AI Overviews or AI Mode. | Keep indexation, crawlability, internal links, visible text, page experience, and accurate structured data as the baseline. | Google also says AIO clicks have been higher quality, but gives no sample, date range, or effect size on this page. |
| [Pew Research Center](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/) | Browsing behavior from 900 U.S. adults: 68,879 Google searches, 12,593 with an AI summary, March 2025. | Treat result clicks, summary-source clicks, session endings, and later branded behavior as separate outcomes. | Observational, Google-only, and search pages were re-run and scraped in April, after the users searched. |
| [Seer’s 2026 AIO study](https://www.seerinteractive.com/insights/aio-impact-on-google-ctr-2026-update) | 53 brands, 5.47M queries, 2.43B organic impressions, and 296.9M paid impressions from January 2025–February 2026. | Segment decisions by intent and citation status. In this cohort, AIO-present organic CTR rose from 1.3% in December 2025 to 2.4% in February 2026; cited AIO queries had 120% more organic clicks per impression than uncited AIO queries. | Directional client data; AIO status was a static recent snapshot applied retrospectively; March 2026 was projected. |
| [GEO: Generative Engine Optimization](https://arxiv.org/abs/2311.09735) | A KDD 2024 benchmark of 10,000 queries. The simulator retrieved five Google sources, generated answers with GPT-3.5, and tested content interventions against internal visibility metrics. | Test clear, sourced, quotable content as an answer-composition hypothesis. | The headline “up to 40%” is not a live citation, traffic, or revenue lift. The source was already in a fixed retrieval set. |

The studies answer different questions.

## What SEO still owns

Google’s own guidance is the least fashionable source in this conversation and one of the most useful. A page must be indexed and eligible for a normal Search snippet before it can appear as a supporting link in AI Overviews or AI Mode. Google also says the features may use query fan-out, which means one user query can lead to several related searches across subtopics and data sources.

That creates a practical SEO decision:

- Fix crawl and indexation problems before buying an AEO platform.
- Make important content available as text, with internal links and a page experience that works for people.
- Keep structured data accurate and matched to visible content.
- Do not build a new “AI file” or special schema block because a vendor checklist says you need one.

None of this guarantees an AI citation. It makes the page eligible for the systems that may retrieve it.

## What AEO adds

AEO becomes useful when the decision is about answer behavior rather than page rank:

| Business question | Useful AEO observation | Unsafe shortcut |
|---|---|---|
| Does the model name us for category questions? | Mention rate by prompt, intent, engine, and competitor. | “Our brand is visible to AI.” |
| Does the model cite our site? | Citation rate, cited URL, cited passage, and source type. | “The model trusts us.” |
| Does a page change improve inclusion? | Repeated pre/post runs with a frozen prompt set and a holdout. | “We added FAQ schema, so citations rose.” |
| Does AI influence demand? | AI referrals, branded search, self-reported discovery, assisted conversion. | “Citation share equals pipeline.” |
| Which content should we create? | Query gaps tied to real buyer prompts and page-level evidence. | A list of generic “best X” prompts. |

Pew’s numbers show why the units need to stay separate. Users clicked a traditional result on 8% of visits when an AI summary appeared, versus 15% without one. They clicked a source inside the summary on only 1% of those visits. That does not mean the cited pages had no influence. It means the click is a narrow outcome and cannot carry the whole measurement story.

Seer’s 2026 update adds another warning. The AIO-present cohort recovered in early 2026 after months of decline. A strategy built around one permanent CTR penalty would now be stale. The useful cut is intent × AIO presence × citation status × time.

## The GEO paper is an optimization study, not a traffic study

The academic GEO paper deserves more respect than most AEO advice gives it, and less than most landing pages give it.

Its benchmark contains 10,000 queries across 25 domains. For each query, the researchers retrieved the top five Google sources, then generated five GPT-3.5 responses at temperature 0.7. The best interventions raised Position-Adjusted Word Count by 41% and Subjective Impression by 28% in the benchmark. Those metrics ask how much of a source’s material appears in the answer and how influential that citation looks under the study’s evaluator.

That is useful evidence that source presentation can affect answer composition. It is not evidence that a live website will receive 41% more citations, 41% more clicks, or 41% more revenue. The source was already inside the five-document context. The experiment changed its share of the answer.

## The operating decision

Use SEO to make content retrievable and useful. Add AEO measurement when the business needs to understand recommendation, citation, source choice, or answer-level perception. Keep the KPIs separate until a real link between them is demonstrated.

Before changing a page, record the prompt set, engine, model/search mode, date, run count, mention definition, citation definition, and downstream outcome. Without that record, a higher score is only a higher score.


## Author disclosure

I cofound Lazyweb and publish Mudpie. This is an owner-written publication, not an independent testing organization. Research notes distinguish observations, sourced reporting and editorial judgment.