Platform index
AI search platforms compared: six engines, six dimensions, one table
Most engine comparisons are decided by user numbers, which is the wrong variable. ChatGPT has roughly 46% of AI assistant usage and Perplexity under 2%, but citation share is a different distribution from traffic share — and the engine with more users is not the engine with more open slots. This page holds the six surfaces that decide AI citations in 2026 against the six dimensions that decide whether your page gets one. Every cell carries a source and a date. Where no public measurement exists, the cell says so rather than guessing.
Last re-dated 2026-10-29. Figures are re-checked whenever a source below publishes a new release.
The matrix
Read across a row to compare engines on one dimension; read down a column to see how one engine behaves on all of them. The two rows that most often get collapsed into a single "AI visibility" score — slots per answer and whether retrieval fires — are the two that change the strategy most.
| Dimension | ChatGPT Search | Perplexity | AI Overviews | AI Mode | Gemini | Claude |
|---|---|---|---|---|---|---|
| Citations per answer | 3.291 | 9.681 | 7.511 | not measured separately4 | 4.971 | 2.411 |
| Answers carrying a citation | 93.9%1 | 90.2%1 | 84.2%1 | not published | 53.1%1 | 60.5%1 |
| Does retrieval fire? | 8 of 20 answers used memory only2 | 19 of 19 answers cited, 10 each2 | always grounded (index-built)13 | grounded by design13 | 62% of answers arrived with no web search3 | no public measurement |
| Cited-URL overlap with peers | 10.2% of cited URLs appear on more than one engine; 17.9% at domain level; 67.4% at brand-name level4 · 11% ChatGPT × Perplexity5 · 8% across five engines6 · AI Mode × AI Overviews 13.7%7 | |||||
| Day-over-day stability | 11.3% cited-set stability8 | ~69% source churn9 | no public churn figure | no public churn figure | no public churn figure | no public churn figure |
| Authority floor / content shape that wins | brand homepages, ranked listicles10 | 59.8% of citations to domains below #100,00011 | 82.91% of branded queries carry one; ~4% of ecommerce queries12 | inline links rising; official-sites test at top13 | names vendors in 60–96% of B2B questions14 | fewest slots, highest bar per slot1 |
Sources
1 Foglift Research five-engine panel, 61,680 answers / 372,111 citations, window 20 Mar–23 Oct 2026, published 24 Oct 2026. · 2 Five-engine retrieval test, 20 fixed questions, 6 Sep 2026. · 3 RanketAI measurement of Korean companies, October 2026 (Gemini app, web search off by default on that query set). · 4 Wellows citation data via TechTimes, 596,723 multi-engine prompts, 1 Sep 2026. · 5 Writesonic, 161,286 prompts, July 2026. · 6 Intender, 3,600 searches / 27,924 citations, September 2026. · 7 Ahrefs, 56 million AI Overviews analysed, February 2026. · 8 Babel42, 660 searches, 12 Oct 2026. · 9 Pineprompt, 18,792 answers across seven engines, 2026. · 10 Kumar et al., arXiv:2606.20065. · 11 Trellner Research TR-2026-009, 7,534 Perplexity citations across 380 buyer-intent categories, 2 Sep 2026. · 12 Ahrefs, 232.8M US desktop results, 29 Sep 2026; BrightEdge Generative Parser, February 2026. · 13 Google Search announcements, 15 Oct 2026. · 14 Chen and Koudas, arXiv:2610.15033, 158,099 answers / 468,037 citations, 10 Oct 2026.
One engine sits outside the six-column table. Grok cites 27 domains per response against Gemini's five (AthenaHQ, State of AI Search 2026, 29 September 2026) — the widest slot count measured on any surface, and the reason Grok citations are the easiest to win and the least durable to hold.
How to read the table without over-reading it
1. Slots are not influence
An answer with ten slots gives you ten ways in; an answer with three gives you three. On that arithmetic Perplexity is the easier door, and two independent studies agree on the direction while disagreeing on the absolute count: 9.68 versus 3.29 in the Foglift panel and 16.35 versus 6.88 in Zhang and Yao's 602-prompt study (21,143 citations, arXiv, 29 April 2026). But ChatGPT exerts the highest influence per cited page of any engine measured — one ChatGPT citation moves brand consideration further than one Perplexity citation. If you count citations to prove progress, Perplexity will always look better than the business result.
2. Retrieval is a gate, not a score
Chen and Koudas found that every page cited had first been returned by that engine's own search, and that retrieval explains 32% of citation variation inside a single site while all fourteen page properties together explain 2.8%. The row above on whether retrieval fires is therefore not a quality signal — it is a statement about whether the dimension applies at all. A ChatGPT programme is partly a retrieval-triggering programme: questions have to be phrased the way someone who does not know the answer phrases them, and the page has to look like the answer rather than the topic.
3. A 10.2% overlap means six programmes, not one
Nine out of ten cited URLs are invisible to the next engine. The gap is not a content-quality gap — it is two separate retrieval systems producing two separate pools. What does carry across is the brand name: 67.4% overlap at the name level against 10.2% at the URL level means the engines agree on who belongs in the answer and disagree on which page proves it. Entity work compounds across engines; page work does not.
Late October 2026: the Google surface moved twice in a week
Both changes affect the same thing — how much of your page survives into the answer. On 15 October 2026 Google announced five changes to links inside AI Overviews and AI Mode: inline links appear far more often and higher in the answer, a test section can show one or two relevant official sites at the top, news carousels now highlight the reader's Preferred Sources across more languages, related links at the end of answers gained thumbnails, and creator names on video results link to Search profiles. Google said Preferred Sources usage has grown more than sevenfold since the start of 2026, without publishing a baseline or any click figures.
Around 21 October, Google began serving AI-generated descriptions inside some AI Overview citation cards — query-conditioned summaries that do not match the page's meta description and may not appear on the page verbatim. The test was spotted by Brodie Clark, covered by Search Engine Roundtable on 23 October, and has no official documentation as of this page's date. The practical consequence is narrow but real: rewriting a meta description will not change card copy, because the input is the passage Google associates with the query, not the meta field.
| What changed | Status | What it means for a publisher |
|---|---|---|
| Inline links more frequent, placed higher | Launching | A citation attached to a specific sentence is now worth more than a card at the bottom. |
| Official-sites section at the top of an answer | Test | Rewards being the canonical, verifiable source for the task — a YMYL and support-doc advantage. |
| AI-generated citation-card descriptions | Observed test (21–23 Oct) | Passages, not meta descriptions, decide card text. Audit the first screen for sentences that stay true alone. |
| External links in AIO answer text | 26.2% daily average, 20 Sep 2026 | Peec AI measurement. Most AIO answers still carry no in-text link at all. |
| Click-through from an AI summary | 1% clicked a link inside the summary | Pew Research Center. Search-result links were clicked in 8% of searches with a summary. |
Which engine first: four situations
The table does not rank engines, because the right first engine depends on what you already have. The four cases below cover most publishers starting from zero, and each links to the page that does the work in detail.
| If you are… | Start with | Because |
|---|---|---|
| A new domain with no brand demand | Perplexity | Widest authority floor (59.8% of citations below rank #100,000) and 9.68 slots per answer. |
| An established brand with documentation | ChatGPT Search | Names known brands in its own queries 67% of the time versus 11% for unknown ones; highest influence per citation. |
| Selling B2B software | Gemini | Names tracked vendors in 60–96% of B2B questions, against ChatGPT's 2–16%. |
| Publishing with measurement budget and no brand | Google surfaces | The largest audience, but AI Mode and AI Overviews share only 13.7% of cited URLs — they are two surfaces, not one. |
For the measured comparison behind the first two rows, see ChatGPT Search vs Perplexity. For why a single "AI visibility" number cannot be aggregated across engines, see citation overlap across engines.
What this table cannot tell you
- ▸ It cannot tell you your own position. Every figure here is an aggregate. Your citation rate on a given query is a separate measurement, and the only one that should drive a decision.
- ▸ It cannot be summed. Slots, inclusion rate and stability are measured on different samples with different methods. Adding them into a composite score invents a number no study produced.
- ▸ It ages faster than a page usually gets edited. Median citation half-life is about 4.5 weeks (Scrunch and Stacker, 3M+ citation events, 26 weeks), and regularly updated sites are cited roughly 3.2 times as often as static ones. A comparison page that is not re-dated is a liability.
- ▸ It cannot substitute for the engine's own retrieval. Google Search results and OpenAI's Responses API together recover only about a quarter of the URLs ChatGPT Web actually cites. Proxies measure proxies.
Frequently asked questions
Which AI search engine should I optimise for first?
Start with the surface that already retrieves you. Chen and Koudas (arXiv:2610.15033, 10 October 2026) found that 100% of cited pages had first been returned by that engine's own search, so a page that is not retrieved cannot be cited no matter how well it is written. Run twenty questions per engine, record which domains come back, and let that decide the order. If no engine retrieves you yet, Perplexity is the fastest diagnostic because it attaches roughly three times as many citations per answer as ChatGPT (9.68 versus 3.29, Foglift Research, 24 October 2026).
Do I need a separate strategy for every engine?
Yes, at the content level. Across 596,723 prompts answered by two or more engines in September 2026, only 10.2% of cited URLs appeared on more than one engine, while brand-name overlap reached 67.4% (Wellows, 1 September 2026). The engines agree on which brands belong in an answer and disagree on which page proves it, so a citation on one engine is not evidence of a citation on another.
Why does ChatGPT sometimes answer without citing anything?
Because ChatGPT is a chat product that chooses whether to search. In a five-engine test on 6 September 2026, 8 of 20 questions were answered from model memory with no sources at all, while Perplexity attached exactly 10 citations to each of its 19 answers. The 93.9% inclusion rate for ChatGPT is computed only over answers that did search, so retrieval has to fire before any citation metric applies.
What changed in Google AI Overviews and AI Mode in October 2026?
Google announced five link changes on 15 October 2026: inline links appear more often and higher in answers, a test section can show one or two official sites at the top of an answer, news carousels highlight Preferred Sources, related links gained thumbnails, and creator names on video results link to Search profiles. Google also said Preferred Sources usage has grown more than sevenfold since the start of 2026. Around 21 October, AI Overview citation cards began showing query-generated descriptions that do not mirror meta descriptions. None of these came with click data.
Is being cited the same as getting traffic?
No. Pew Research Center measured that 1% of searches with an AI summary produced a click on a link inside the summary, versus 8% clicking a search-result link. Peec AI put the share of AI Overviews containing an external link in the answer text at a 26.2% daily average on 20 September 2026. Citation and click have to be measured as separate outcomes, on separate dashboards.
Where do the engines disagree most?
On the authority floor. Trellner Research (TR-2026-009, 2 September 2026) found 59.8% of Perplexity citations in software buyer-intent categories pointed to domains ranked worse than #100,000 globally, while ChatGPT leans toward brand homepages and ranked listicles (Kumar et al., arXiv:2606.20065). A new publisher can enter through Perplexity's retrieval pool and still be invisible to ChatGPT, and the reverse is also true.
Why is there no number in some cells?
Because no public measurement exists for that engine and dimension, and a plausible-looking estimate is worse than a blank. Claude has published citation-density and inclusion figures but no day-over-day churn study; AI Mode has never been measured separately from AI Overviews at URL level, and the two surfaces share only 13.7% of cited URLs despite 86% semantic similarity (Ahrefs, February 2026).
How often should the numbers on this page be re-checked?
More often than most teams expect. A Scrunch and Stacker analysis of more than three million citation events across eight industries and 26 weeks put the median citation half-life at about 4.5 weeks — roughly 3.4 weeks on ChatGPT and close to 6 on Perplexity — and found regularly updated sites cited about 3.2 times as often as sites that are not updated. This page is re-dated whenever a source in it changes.
About the publisher. GeoAura is an English-language research publisher covering Generative Engine Optimization and AI citation measurement. It operates one site, geoaura.world, and measures how ChatGPT Search, Perplexity, Google AI Overviews, AI Mode, Gemini and Claude cite sources — publishing the method and the raw numbers behind every claim, including the ones that contradict its own earlier positions. GeoAura is not affiliated with geoaurora.cn, aurasearch.ai, the GenAura brand, or Aurora Intelligence.
Corrections: figures on this page are withdrawn rather than quietly overwritten. The log, including the four-state quotation standard, is at /corrections.