Key findings
- Answering the query directly is the most robust signal. Per level, the chance of a citation rises about 1.6-fold in the Google AI Overview and about 2.5-fold in Perplexity, at the same Google ranking.
- Answer-first, meaning sections that start with the answer, is also linked with citations (2.1 and 2.7 times). The automatic rating measures this feature only roughly, though.
- Only in Perplexity are sources cited in the text (1.8 times) and FAQ schema (1.9 times) also linked with citations. For FAQ schema, a before-and-after measurement argues against a cause.
- Lists, tables, numbers, short sentences and external links show no significant link in any service.
- A typical GEO score is barely better than chance: it hits in about 53% of cases. Google ranking hits in about 77%.
Five common assumptions and what our data shows
Four of five common GEO assumptions do not hold up in our measurement, the fifth only in one service. All values come from this data test and apply at the same Google ranking. 1× means the same chance of a citation with or without the feature.
| Assumption | Google AI Overview | Perplexity | What our data shows |
|---|---|---|---|
| More lists mean more AI citations | 0.98× (0.91 to 1.05) | 0.98× (0.92 to 1.05) | No significant link, not in ChatGPT either |
| Numbers and statistics mean more citations | 0.98× (0.95 to 1.00) | 1.00× (0.98 to 1.01) | No significant link |
| Short sentences mean more citations | 0.99× (0.94 to 1.03) | 1.00× (0.98 to 1.02) | No significant link |
| FAQ schema means more citations | 1.40× (1.00 to 1.97) | 1.87× (1.35 to 2.60) | Significant only in Perplexity, not in the AI Overview |
| A high GEO score predicts citations | hits in 53% | hits in 56% | Barely better than chance (50%), Google ranking hits in 77% |
Odds ratio with 95% range. For the GEO score: share of pairs in which the cited page has the higher score. Significant means it holds after correction for 16 simultaneous tests.
What does work: pages that answer the query directly. The next section shows this.
Why we tested GEO recommendations
Most GEO recommendations are rarely tested against the AI answers people actually see. Typical checklists recommend short sentences, lists, statistics, sources, author details and a summary at the top. Often these are added up into a score.
We wanted to know whether pages that meet a recommendation are linked as a source more often, compared with pages that ranked similarly on Google for the same query. As our analysis of AI answers shows, Google ranking is the strongest single factor. Without controlling for it you only measure that well-ranking pages are cited more often.
What really goes with AI citations
The most robust link is whether a page answers the query directly. It is the only feature that stays significant in both the Google AI Overview and Perplexity when the main other features are considered at the same time (AI Overview 1.48 times, Perplexity 2.23 times).
Content, trust and technical: chance of a citation at the same Google ranking
Odds ratio with 95% range. Right of the dashed line means cited more often. Filled dot: holds after correction for 16 simultaneous tests.
Values as a table
| Recommendation | Google AI Overview | Perplexity | ChatGPT |
|---|---|---|---|
| Page answers the query directly | 1.64 (1.25 to 2.15) significant | 2.48 (1.89 to 3.26) significant | 1.10 (0.68 to 1.79) |
| Sections start with the answer | 2.10 (1.32 to 3.34) significant | 2.73 (1.75 to 4.28) significant | 1.47 (0.64 to 3.40) |
| Headings phrased as questions | 2.20 (1.20 to 4.02) | 2.27 (1.22 to 4.23) | 0.63 (0.16 to 2.44) |
| Summary at the top | 1.47 (1.06 to 2.04) | 1.13 (0.83 to 1.55) | 0.70 (0.40 to 1.24) |
| Sections understandable on their own | 1.62 (1.10 to 2.38) | 1.22 (0.86 to 1.75) | 1.50 (0.80 to 2.82) |
| Intro gets to the point | 1.11 (0.95 to 1.30) | 1.21 (1.03 to 1.41) | 1.07 (0.81 to 1.42) |
| Sources cited in the text | 1.05 (0.77 to 1.44) | 1.75 (1.28 to 2.39) significant | 0.88 (0.48 to 1.63) |
| Author or expertise named | 0.87 (0.68 to 1.12) | 1.31 (1.02 to 1.68) | 1.08 (0.70 to 1.69) |
| Updated in the last year | 1.91 (1.21 to 3.00) | 1.50 (0.96 to 2.34) | 0.58 (0.11 to 3.12) |
| FAQ schema (structured data) | 1.40 (1.00 to 1.97) | 1.87 (1.35 to 2.60) significant | 1.85 (1.10 to 3.10) |
Closely related is answer-first, meaning sections start with the answer. On its own this link is also significant; in the joint model it is largely absorbed by answering directly. In practice it is the same principle: the page delivers the answer instead of leading up to it. A section “How much does X cost?” that starts with the price does better than one that starts with “Costs depend on many factors”.
What only matters in Perplexity
Perplexity cites pages with sources in the text and with FAQ schema more often, the Google AI Overview does not. With sources the chance in Perplexity is about 1.8 times as high, with FAQ schema about 1.9 times. Both remain in the joint model. In the AI Overview the value for FAQ schema is 1.40 and for sources 1.05, neither significant.
Sources are different: the Princeton lab experiment, which deliberately changed texts, also shows an effect.
Do lists, tables and statistics help?
Not measurably in our data, and if at all, only very little. More lists, tables, numbers, shorter sentences and more external links show no significant link with citations in any of the three services.
Format rules: no significant link
Same chart type. All dots lie close to 1, meaning the same chance with or without the feature.
Values as a table
| Recommendation | Google AI Overview | Perplexity | ChatGPT |
|---|---|---|---|
| More lists | 0.98 (0.91 to 1.05) | 0.98 (0.92 to 1.05) | 0.85 (0.74 to 0.98) |
| Has a table | 1.08 (0.80 to 1.46) | 1.08 (0.82 to 1.42) | 1.56 (0.98 to 2.47) |
| More numbers and statistics | 0.98 (0.95 to 1.00) | 1.00 (0.98 to 1.01) | 0.99 (0.96 to 1.02) |
| Shorter sentences (per 5 words) | 0.99 (0.94 to 1.03) | 1.00 (0.98 to 1.02) | 1.04 (0.97 to 1.12) |
| More external links | 1.05 (0.96 to 1.15) | 1.00 (0.95 to 1.05) | 1.00 (0.94 to 1.07) |
| Longer text | 1.09 (1.01 to 1.17) | 1.22 (1.12 to 1.33) significant | 0.98 (0.86 to 1.12) |
A pre-registered study with about 114,000 observations finds effects of 3 to 4% per standard deviation for statistics and lists (Lee 2026). Both measurements reach the same conclusion: these elements barely separate cited from non-cited pages. That does not mean they do not help readers. But when reworking a page for AI answers, the effort is better spent on the direct answer first.
Do you need llms.txt or special schema?
For Google, no. In its official documentation on AI features, Google says there are no additional requirements for AI Overviews and AI Mode. New machine-readable files, AI text files or special schema.org markup are not needed. The same fundamentals as for regular search apply.
We have not measured llms.txt ourselves. Our FAQ schema data fits Google’s statement, though: in the AI Overview it shows no significant link with citations. Time is better invested in content that answers the question directly.
How reliable are GEO scores?
A typical total score made of common recommendations barely predicts which page gets cited. We weighted the features using a widespread 100-point scheme and compared every cited page with every non-cited page for the same query.
Which signal predicts which of two pages is cited?
For each query, every cited page was compared with every non-cited page. The bar shows how often the cited page had the better value. The vertical mark is 50%, pure chance.
Google AI Overview
Perplexity
In the Google AI Overview the cited page has the higher score in only about 53% of pairs, in Perplexity about 56%. The score mixes a few features with a significant link with many without and dilutes the signal. A single feature such as answering directly hits better than the whole score.
Rating twice is no proof of quality. Content features can only be rated at scale by a language model. Only a comparison with an independent human judgement is meaningful. So a person rated 60 randomly drawn pages without knowing the model’s rating:
For answering directly, agreement is good enough for comparing groups of pages, not for judging a single page. For answer-first the extract often contained navigation or footers instead of the actual text. We therefore treat this finding as an indication only.
What matters more: Google ranking or page structure?
In the Google AI Overview it is ranking, in Perplexity both are about equal within the top five positions. Ranking hits in about 77% of pairs in the AI Overview, answering directly in about 58%. In Perplexity both are at about 64% and 62%.
A page outside the top Google results is rarely cited in the AI Overview, however well it is built. If it ranks high, structure helps decide whether it or a neighbouring page is cited. In Perplexity a well-answering page in position 4 or 5 has good chances against a weaker page in position 1. Our measurement says nothing about larger gaps.
What other studies show
The big picture is backed by several sources: Google ranking is the strongest factor, the answer at the top helps, format rules barely help. Our data confirms this for German-language queries.
- Lee (2026): about 114,000 pages, pre-registered. Positions 1 to 3 have 7.8 times the chance of positions 11 to 30. Answer first 1.09, lists 1.04, statistics 1.03 per standard deviation.
- Ahrefs (2026): 1,885 pages that added structured data vs 4,000 unchanged pages. No rise in citations.
- Wu et al. (2025): the AI Overview cites text that is easy for language models to predict more often.
- Indig (2026): about 44% of ChatGPT citations come from the first third of a page.
- Aggarwal et al., Princeton (2024): deliberately rewritten texts with quotes, sources and statistics gain up to about 40% visibility in a lab setting.
New in our analysis: we also rate whether the content answers the query, not just format features. We measure the Google AI Overview with German-language queries. And we re-measure the same pages until December 2026 to test whether better structure really brings more citations. The analysis plan was fixed in advance.
And ChatGPT?
Our data allows no statement for ChatGPT. ChatGPT links sources less often, so only 143 queries with 623 pages were usable. The ranges are so wide that even clear effects remain possible. “No finding” here means “not measurable”, not “does not work”. Our guide to ChatGPT describes which factors are discussed.
How to optimise your website for AI search
Check ranking first, then the answer, and format last.
- 01Check your rankingPages outside the top Google results are rarely cited in the AI Overview. Where that is the bottleneck, classic SEO helps more than any rewording.
- 02Answer the query itselfFor your key queries, check whether the page answers the question or only covers the topic. The answer belongs in the first sentence of the matching section (answer-first).
- 03Cite sources for PerplexityTraceable sources in the text are the measure with the best evidence for Perplexity: a link in our measurement and an effect in the lab experiment.
- 04Do not just work through checklistsLists, tables and numbers are good for readers but bring hardly any extra citations according to this data. The same applies to FAQ schema, llms.txt and a high GEO score.
How we measured
We compared unchanged pages that ranked in the top Google positions for the same query and checked which of them the AI answer links as a source.
- Queries: random selection of queries from Google Search Console, asked by people, focus on Austria, three words or more, the same basis as our analysis of AI answers. Only queries that occur on at least three websites, without company, brand or personal names.
- Pages: for every query the websites in positions 1 to 5 of regular Google Search at the same moment. Cited means: exactly this page is linked as a source in the AI answer.
- Features: countable ones (lists, tables, sentence length, numbers, links, date, structured data, text length) extracted automatically from the main text. Content features rated by a language model on a 0 to 2 scale, based on title, intro and the first eight sections. A person rated 60 pages independently.
- Comparison: conditional logistic regression per query with Google ranking as a control. Analysed are queries with at least one cited and one non-cited page: AI Overview 388 queries with 1,624 pages, Perplexity 380 with 1,592, ChatGPT 143 with 623.
- Statistics: p-values Holm-adjusted for 16 tests per service. Hit rates in pairwise comparison with 95% bootstrap over queries.
Data download
All values of this analysis are available as CSV files.
geo-empfehlungen-je-dienst.csvHit rate in pairwise comparisonShare of pairs (cited vs not cited page for the same query) in which the cited page has the better value, per service.geo-trefferquote-paare.csvLicence: CC BY 4.0. Please name “Atziry” as the source with a link to this page. Aggregated values only, no queries and no websites.
