Answer engines
PerplexityBuilt to answer every question from a web search, with numbered sources beside the answer. In our API test Perplexity searched on all 240 of its answers.
Use one when you need to check where each claim came from.Buyer's guide
The 9 Best AI Search Engines in 2026, Compared
The best AI search engines in 2026, compared with Prefer's test of 960 AI answers: how often Perplexity, ChatGPT, Gemini, Claude and Google AI search the web.
Also inالعربيةPortuguês (Brasil)
The short answer
Three picks, for three situations.
What is the best AI search engine in 2026?
Perplexity, if you want to check where an answer came from: in Prefer's September 2026 API test of 960 AI answers it showed sources on every answer and gave almost the same sources when asked twice. Those figures come from its sonar-pro API model, and Perplexity lists more citations per answer as a Pro benefit, so use Pro for research. Google AI Mode is second and already inside Google Search. ChatGPT is the best everyday assistant, and its search works without signing in.
Best for cited research (Pro) 
Perplexity Sources on every answer, same sources twice
Perplexity Sources on every answer, same sources twice Best inside Google Search 
Google AI Mode Already in the search engine you use
Google AI Mode Already in the search engine you use Best everyday assistant 
ChatGPT Free web search, even signed out
DisclosurePrefer sells software that measures how AI engines cite brands, so we have a stake in AI search. None of these engines is our product, and none paid to be listed. Our numbers come from API answers and Google captures collected on 13 September 2026, not the consumer apps, and the full dataset is public.
In short
The takeaways, up front.
- In our September 2026 API test, Perplexity showed sources on all 240 of its answers and cited 19.5 per answer. Gemini cited 8.6, Claude 4.6 and ChatGPT 3.05.
- Asked the same question twice, Perplexity's two runs shared, on average, 96% of the sites cited across both. Claude's shared 68%, ChatGPT's 37% and Gemini's 35%.
- The engines mostly read different sites, but agree on the big ones: of 1,329 sites cited, 966 were cited by one engine only, yet 23 of the 25 most-cited sites were cited by three or four.
- All nine engines are free to use, per each maker's own pages read on 19 September 2026. Paid plans add usage and extras, such as more citations per answer on Perplexity Pro.
- Copilot, Grok, Meta AI and DeepSeek were not in our test. They rank after the measured engines, on their makers' documentation, dated on this page.
How we evaluated
Four criteria. Nothing else counted.
Perplexity is the best AI search engine for research you need to check, per Prefer's September 2026 API test of 960 AI answers: it showed sources on every answer. Google AI Mode is second and already inside Google Search, and ChatGPT is the best everyday assistant. We compared nine engines using that test and each maker's own pages. All nine are free to use.
01
Does it show a source on its answers?
An answer built from a live search shows its sources and can be checked. An answer from memory cannot. We counted answers that showed at least one source: Perplexity 240 of 240, Google AI Mode 74 of 80, ChatGPT and Gemini 213 of 240 each, Claude 192 of 240 (13 September 2026). This is the first sort key.
02
Does it give the same sources twice?
We asked every question three times. Two runs shared, on average, 96% of their sites on Perplexity, 68% on Claude, 37% on ChatGPT and 35% on Gemini. This is the second sort key. Google AI Mode was captured once per question, so it has no score here.
03
What does its maker document?
For Copilot, Grok, Meta AI and DeepSeek, which we did not measure, we read each maker's own pages on 19 September 2026: does it search the web, does it cite, and what caveats does the maker state. Measured engines rank above unmeasured ones.
04
Free access and reach
All nine are free to use. Where engines tie, the one that is easier to reach ranks higher: no sign-in needed, or built into apps people already use. Sources per answer is reported, not scored, because twenty links are not better than three you will open.
Three kinds of engine
They are not all the same product.
Assistants that search
ChatGPT, Gemini, Claude, Copilot, Grok, Meta AI, DeepSeekGeneral assistants that decide question by question whether to search, then weave the sources into a reply. The three we measured searched on 80% to 90% of answers and cited far fewer sources than Perplexity.
Use one when you want a capable assistant that also looks things up.Search engines with AI answers
Google AI Mode and AI OverviewsGoogle Search with an AI summary on top (AI Overviews) or as a conversation (AI Mode), both drawing on Google's index. An AI Overview appeared on 75 of our 80 questions.
Use one when you already start on Google and want an answer plus links.The ranked list
Every tool worth shortlisting.
Measured engines first, by share of answers showing a source, then same sources when asked twice. Unmeasured engines after, by what their makers document, then reach.
01
perplexity.ai
Perplexity
perplexity.ai Perplexity searches the web for every question and answers with numbered sources. In our September 2026 API test (its sonar-pro model) it showed sources on all 240 answers, cited more sources than any other engine, and gave almost the same sources when asked the same question twice.
Strengths
Showed sources on 240 of 240 answers in our API test and never cited fewer than 11
The steadiest engine we measured: two runs of the same question shared, on average, 96% of the sites cited across both
Cited 748 different sites, more than any other engine, with Reddit in 98 of 240 answers and LinkedIn in 73
Limitations
Our numbers come from its sonar-pro API model, not the free app. Its help center lists '10x as many citations per answer' as a Pro benefit, so the free plan may show fewer
With about 20 sources per answer, no single link carries much weight, and the checking is on you
VerdictThe one to use when you want to see, and check, where an answer came from. For the full source lists we measured, use Pro. Visit site →
02
google.com
Google AI Mode
google.com AI Mode splits your question into subtopics and searches each one, Google says. In our capture of the same 80 questions it showed sources on 74 of them (92.5%), with 5.8 sources per answer, and an AI Overview appeared on 75 questions with 8.6 sources each.
Strengths
Built into Google Search: Google says AI Mode passed one billion monthly users in May 2026
Showed sources on 74 of our 80 questions, a higher share than ChatGPT or Gemini in our API test (Google was captured once per question, a different method)
Google says AI Mode divides your question into subtopics and searches for each one at the same time
Limitations
The two surfaces cite different pages: only 32% of AI Mode's sources also appeared in the AI Overview for the same question
Captured once per question, so we could not measure whether it repeats its sources; YouTube led its sources (51 of 76 answers), so expect video links as often as articles
VerdictThe AI search engine you already have. Good for everyday questions; check the links, because the Overview and AI Mode often cite different pages. Visit site →
03
chatgpt.com
ChatGPT
chatgpt.com ChatGPT decides question by question whether to search. With search switched on in our API test it showed sources on 213 of 240 answers, but cited only 3.05 per answer, the fewest of the four chat engines we measured.
Strengths
OpenAI's help center lists web search on every plan, Free included, and says people who are not signed in can use it too (read 19 September 2026)
Searched on 216 of 240 answers in our API test, and on all three runs for 70 of 80 questions
Short source lists, a median of 3 per answer, so every link it shows is quick to check
Limitations
Two runs of the same question shared, on average, only 37% of the sites cited across both
27 of its 240 answers cited no source: 24 that skipped search and 3 that searched
OpenAI's own help page says search results and citations 'can be incomplete, outdated, or incorrect'
VerdictThe easy everyday pick: free, no sign-in needed, and it searches when it matters. Open the links, because there are only a few. Visit site →
04
gemini.google.com
Gemini
gemini.google.com In our API test Gemini showed sources on 213 of 240 answers, the same as ChatGPT, and cited 8.6 per answer, second only to Perplexity. It was the only chat engine that cited YouTube, in 113 of its 240 answers.
Strengths
Second-most sources in our API test: 8.6 per answer and 580 different sites across 240 answers
The only chat engine that cited YouTube (113 of 240 answers), plus Reddit in 93
Google's plans page lists Deep Research in the free version (read 19 September 2026)
Limitations
Least steady: two runs of the same question shared, on average, 35% of their sites, and only 2 of 200 run pairs cited identical sites
Skipped search on 27 of 240 answers, and those 27 cited nothing
VerdictGood when you want Google's index plus video. Ask twice before you trust its source list. Visit site →
05
claude.ai
Claude
claude.ai In our API test Claude showed sources on 192 of 240 answers (80%), the fewest of the four chat engines, and cited 4.6 per answer. After Perplexity, its sources were the steadiest between runs.
Strengths
Anthropic's help center says 'every response includes citations' when Claude searches, and free accounts can search
Two runs of the same question shared, on average, 68% of their sites, second only to Perplexity
Answers that searched cited a median of 6 sources, a list short enough to open
Limitations
Searched least of the four: 48 of its 240 answers did not search and cited nothing
On some accounts web search is a switch in the '+' menu, and Team and Enterprise owners must enable it first
Cited no Reddit and no YouTube in 240 answers on our AI search and marketing questions
VerdictA good pick when you want a tight source list that holds up when you ask again. Visit site →
06
copilot.microsoft.com
Microsoft Copilot
copilot.microsoft.com We did not measure Copilot. Its place rests on Microsoft's pages: since February 2026 Bing Webmaster Tools reports how often Copilot cites a site's pages, and Microsoft 365 plans build it into Office apps with higher usage than free.
Strengths
Cites web pages in its answers: since February 2026, Bing Webmaster Tools shows site owners how often Copilot cites them (Bing, 10 February 2026)
Microsoft 365 plans list Word, Excel, PowerPoint and Outlook 'with Microsoft Copilot' and 'higher usage than free' (Microsoft, read 19 September 2026)
Limitations
Not in our study, so we have no measured search rate or source count for it
We have no dated source on how the free Copilot app picks its sources; the Bing grounding docs we found cover Copilot Studio agents
VerdictThe sensible pick if your day already runs on Microsoft apps. Visit site →
07
grok.com
Grok
grok.com We did not measure Grok. Its access to live X posts is what sets it apart, per xAI and X's own help center, and X's help center also states the clearest accuracy caveat of any maker here.
Strengths
xAI says it searches the web and X live, with 'live citations from primary sources across the web' (read 19 September 2026)
Can search public X posts by keyword, meaning, user and thread, per xAI's developer docs
Free to try on the web and in the apps; SuperGrok adds higher limits and multi-agent reasoning
Limitations
X's help center calls it 'an early version' that 'may confidently provide factually incorrect information'
Public X posts are unvetted, so a live answer can repeat whatever is trending
Not in our study, so we have no measured search rate or source count for it
VerdictReach for it when a story is breaking on X, and check it before you repeat it. Visit site →
08
meta.ai
Meta AI
meta.ai We did not measure Meta AI. Meta says it can search across the web, and since June 2026 Facebook's AI Mode answers from public posts in Groups and Reels.
Strengths
Meta says people use it across WhatsApp, Instagram, Facebook and Messenger, and that it can search across the web (29 April 2025)
Facebook's AI Mode (15 June 2026) answers from what people say publicly in Groups and Reels, useful for opinions and experiences
Meta says Meta AI stays free for everyday use; Meta One bundles add more usage (15 September 2026)
Limitations
On Facebook, its AI Mode draws on public posts in Meta's apps rather than the wider web
Not in our study, and we have no dated source on how it shows its sources
VerdictConvenient where you already chat. Not the tool for research you need to verify. Visit site →
09
deepseek.com
DeepSeek
deepseek.com We did not measure DeepSeek. It is here for its license: the DeepSeek-V4-Pro model card says the model weights are under the MIT License.
Strengths
The DeepSeek-V4-Pro model card says the repository and model weights are licensed under the MIT License (read 19 September 2026)
Its privacy policy says the app uses third-party search APIs to answer, so it can look things up
Limitations
Its privacy policy says it collects, processes and stores personal data in the People's Republic of China
V4-Pro has 1.6 trillion parameters, so running it yourself needs serious hardware
Not in our study, and we have no dated source on how its app shows sources
VerdictThe pick when open weights matter more than a polished search experience. Visit site →
Side by side
The whole field, one table.
Best-for, the standout strength, and the main limitation for each tool.
| Tool | Best for | Strength | Limitation | From |
|---|---|---|---|---|
| Perplexity | Cited research (Pro) | Most sources in our API test | Free plan may show fewer | Free; Pro and Max plans |
| Google AI Mode | Inside Google Search | Built into Google Search | Overview and AI Mode cite different pages | Free in Google Search |
| ChatGPT | Everyday assistant | Search works signed out | About 3 sources per API answer | Free, even signed out; paid plans |
| Gemini | Video and community sources | Cited YouTube most (API test) | Sources shift between runs | Free; Google AI Plus, Pro, Ultra |
| Claude | Steady, short source lists | Steady sources in our API test | Searched least (80%) | Free; paid plans |
| Microsoft Copilot | Free, in Microsoft 365 | Cites pages; in Office apps | Not in our study | Free; Microsoft 365 plans |
| Grok | Live X posts | Searches the web and X live | Not in our study | Free; SuperGrok plan |
| Meta AI | Inside Meta's apps | Built into Meta's apps | Not in our study | Free; Meta One plans |
| DeepSeek | Open weights | MIT-licensed model weights | App data stored in China | Free; open weights (MIT) |
What changed in 2026
Four shifts that reshaped the category.
This list looked different a year ago. These are the changes that moved it.
01
Web search became free, even signed out
OpenAI's help center lists web search on every ChatGPT plan, Free included, and says people who are not signed in can use it too. Claude's free accounts can search, and Google lists Deep Research in Gemini's free version (all read 19 September 2026). Price stopped being the gate.
02
Google joined AI Overviews and AI Mode
By May 2026 Google had joined the two: a follow-up to an AI Overview continues in AI Mode, live worldwide, and Google said AI Mode had passed one billion monthly users. For most people, Google is now an AI search engine too.
03
Search went free; paid plans add volume and extras
Paid plans now sell more than access: Perplexity Pro lists '10x as many citations per answer', SuperGrok adds higher limits and multi-agent reasoning, Microsoft 365 gives Copilot higher usage than free, Google AI plans raise Gemini's limits, and Meta One adds Meta AI usage while Meta AI stays free for everyday use.
04
The engines split the long tail of the web
Of 1,329 sites cited in our test, 966 were cited by only one of the four chat engines, though 338 of those appeared in a single answer. The big sites are shared: 23 of the 25 most-cited sites were cited by three or four engines.
The basics
What is an AI search engine?
An AI search engine answers your question in plain language and links the web pages it used. Some search on every question (Perplexity). Most assistants decide question by question (ChatGPT, Gemini, Claude). Google adds AI answers on top of ordinary search (AI Overviews and AI Mode).
The difference from a plain chatbot is the search step. An answer built from a live search can be checked against its sources. In our test, no answer that skipped search cited anything.
Glossary6 terms
- AI search engine
- A service that answers a question in plain language from a web search and links the pages it used.
- Grounding
- Basing an answer on pages fetched when you ask, not on the model's memory. Google describes grounding as connecting Gemini to real-time web content.
- Citation
- A link to a page the answer used. Perplexity numbers its sources; ChatGPT and Claude link a few per answer.
- AI Overview
- Google's AI summary at the top of an ordinary results page. It appeared on 75 of our 80 questions.
- AI Mode
- Google's conversational AI search. Google says it divides a question into subtopics and searches for each one.
- Query fan-out
- Running several related searches for one question. Google says AI Overviews and AI Mode may use it.
An AI search engine answers your question in plain language and links the web pages it used, so you can check the answer instead of sorting through a page of links. Perplexity is the best AI search engine in 2026 for research you need to verify, based on Prefer’s September 2026 test of 960 AI answers; Google AI Mode is second and already inside Google Search, and ChatGPT is the best everyday assistant. This list compares nine engines for the people who use them. If you want software that tracks how these engines mention your brand, see the best AI visibility tools or the best AI search monitoring tools instead.
How we compared nine AI search engines#
We did two things:
- We read each maker’s own pages on 19 September 2026 for what the engine does, where its results come from and what is free. Every source is dated in the list at the foot of this page.
- We measured five of the engines. On 13 September 2026 we asked ChatGPT, Perplexity, Gemini and Claude the same 80 questions through their APIs, three times each, with web search switched on: 960 answers in all. For the same 80 questions we captured Google’s AI Overview and AI Mode once each. The method and the full dataset are in the September 2026 AI answer study.
Three limits matter. The chat answers came through each engine’s API (models gpt-5.6-luna, sonar-pro, gemini-3.8-flash and claude-sonnet-5), not the free apps, so the apps may behave differently. All 80 questions were about AI search and marketing, so the sites each engine cited reflect that topic. And we did not grade whether the answers were right. Microsoft Copilot, Grok, Meta AI and DeepSeek were not in the test; their places rest on their makers’ documentation.
The four chat engines in our API test, side by side#
| Engine | Searched the web | Sources per answer | Different sites cited | Answers with no source | Two runs shared, on average |
|---|---|---|---|---|---|
| Perplexity | 240 of 240 (100%) | 19.48 | 748 | 0 | 96% of sites |
| Gemini | 213 of 240 (88.8%) | 8.64 | 580 | 27 | 35% of sites |
| Claude | 192 of 240 (80%) | 4.58 | 333 | 48 | 68% of sites |
| ChatGPT | 216 of 240 (90%) | 3.05 | 218 | 27 | 37% of sites |
Measured 13 September 2026 through each engine’s API, 240 answers per engine. No answer that skipped search cited a source. “Two runs shared” is the share of all the sites cited across two runs of the same question that both runs cited, averaged over the run pairs where both cited at least one site.
In our API test, Perplexity cited about six times as many sources as ChatGPT#
Perplexity cited 19.5 sources per answer in our API test, about six times ChatGPT’s 3.05. Gemini cited 8.6 and Claude 4.6. Perplexity never cited fewer than 11 sources on an answer. ChatGPT’s median was 3, and Claude’s median on answers that searched was 6.
That makes the choice a trade. A long list lets you check a claim against several pages, but no single link carries much weight and the checking is on you. A short list is quicker to open, and each link matters more. One caution on Perplexity: our numbers come from its sonar-pro API model, and Perplexity’s help center lists “10x as many citations per answer” as a Pro benefit, so the free plan may show fewer.
Whichever engine you use, three steps make an answer checkable:
- Open two of the cited sources, not just the first one.
- Find the sentence in each that supports the claim. If neither says it, treat the claim as unsupported.
- Ask the same question again. If the sources change completely, the first answer was a sample, not a verdict.
Search rates tell the same story as the source counts. Perplexity searched on every answer. ChatGPT searched on 216 of 240 and Gemini on 213. Claude searched least, on 192 of 240 (80%), and the 48 answers where it did not search cited nothing. Google’s grounding documentation explains how this happens with Gemini: the model “determines if a Google Search can improve the answer” before it searches.
In our API test, only Perplexity gave the same sources when asked twice#
We asked every question three times, and the three answers came back within 23 seconds of each other. On Perplexity, two runs of the same question shared, on average, 96% of the sites cited across both. On Claude it was 68%, on ChatGPT 37% and on Gemini 35%. Perplexity cited an identical set of sites on 199 of 240 pairs of runs, Claude on 55 of 179, ChatGPT on 27 of 206 and Gemini on just 2 of 200.
For a reader, this is a practical test of trust. If an engine cites different pages each time you ask, one answer is a sample, not a verdict. Ask ChatGPT or Gemini twice before you rely on a source list. More in do AI answers change when you ask twice?
In our test, each engine leaned on a different part of the web#
Of the 1,329 sites the four chat engines cited, 966 (72.7%) were cited by only one engine, and just 29 by all four. On 69 of our 80 questions, no single site was cited by all four. That split is mostly in the long tail: 338 of the 966 single-engine sites appeared in just one answer, and 23 of the 25 most-cited sites were cited by three or four engines. The engines agree on the big sites and differ on everything below them.
The platforms they lean on differ sharply:
| Answers citing it (of 240) | YouTube | ||
|---|---|---|---|
| Perplexity | 98 | 0 | 73 |
| Gemini | 93 | 113 | 0 |
| ChatGPT | 16 | 0 | 0 |
| Claude | 0 | 0 | 2 |
Gemini was the only chat engine that cited YouTube, in 47% of its answers. Perplexity was the only one that cited LinkedIn in any volume, and 55 of its 73 LinkedIn answers cited LinkedIn articles (Pulse). Claude cited no Reddit and no YouTube at all.
Each engine’s most-cited sites show the same split:
| Engine | Most-cited sites (answers, of 240) |
|---|---|
| ChatGPT | Google Search Central 58, arXiv 44, TechRadar 27 |
| Perplexity | Reddit 98, Search Engine Land 87, LinkedIn 73 |
| Gemini | YouTube 113, Reddit 93, Semrush 39 |
| Claude | Profound 26, SE Ranking 17, Ahrefs 16 |
Two cautions. Profound, SE Ranking, Semrush and Ahrefs sell SEO or AI visibility software, the same kind of product Prefer sells, and 6 of Claude’s 26 Profound answers came from questions that name Profound. And the topic moves the list: ChatGPT cited Reddit in only 16 of 240 answers here, yet in our August 2026 Reddit study of software buyer questions, measured with a different method, Reddit was its top source on 25 of 25 queries. What an engine reads depends on what you ask. For ChatGPT, see which sources ChatGPT cites most.
The engines also differ in how much of their reading list is theirs alone. 136 of the 218 sites ChatGPT cited (62%) were cited by no other engine, against 424 of 748 for Perplexity (57%), 310 of 580 for Gemini (53%) and 96 of 333 for Claude (29%). Claude mostly cited sites the others cited too.
On our 80 questions, Google’s AI Overviews and AI Mode cited different pages#
Google runs two AI surfaces on one index. An AI Overview appeared on 75 of our 80 questions and cited 8.6 sources on average. AI Mode returned an answer for 76 of 80, showed sources on 74, and cited 5.8 per answer. YouTube was the most-cited site on both.
| Google surface | Answered | Sources per answer | Most-cited sites (questions) |
|---|---|---|---|
| Google AI Overviews | 75 of 80 | 8.64 | YouTube 38, Reddit 29, LinkedIn 18 |
| Google AI Mode | 76 of 80 | 5.80 | YouTube 51, Reddit 26, LinkedIn 14 |
They do not cite the same pages. On the 69 questions where both cited sources, only 32% of AI Mode’s sources also appeared in the AI Overview. Google now lets a follow-up question to an AI Overview continue in AI Mode (live worldwide since May 2026, Google says), so one search can show you two different reading lists. The full comparison is in AI Overviews vs AI Mode.
What we know about the four engines we did not measure#
These four rank after the measured engines, on their makers’ documentation, read on the dates in the source list. We have no measured search rate or source count for any of them.
- Microsoft Copilot cites web pages in its answers: since February 2026, Bing Webmaster Tools shows site owners how often Copilot cites their pages. Microsoft 365 plans list Word, Excel, PowerPoint and Outlook “with Microsoft Copilot” and “higher usage than free”. The Bing grounding documentation we found covers Copilot Studio agents, not the free app.
- Grok decides whether to search public X posts and run a live web search, per X’s help center, and xAI’s product page lists “live citations from primary sources across the web”. The same help page calls it “an early version” that “may confidently provide factually incorrect information”.
- Meta AI is built into WhatsApp, Instagram, Facebook and Messenger and can search across the web, per Meta. On Facebook, its AI Mode (June 2026) answers from what people post publicly in Groups and Reels.
- DeepSeek publishes its models as open weights: the DeepSeek-V4-Pro model card says the model weights are licensed under the MIT License. Its privacy policy says the app uses third-party search APIs and stores personal data in the People’s Republic of China.
How we ranked them, and what we left out#
Measured engines rank first, sorted by two numbers: the share of answers that showed at least one source, then how often two runs of the same question cited the same sites. Sources per answer is reported, not scored, because twenty links are not better than three you will open.
- Perplexity showed sources on 240 of 240 answers and its runs shared 96% of sites, so it leads on both. Its figures come from the sonar-pro API model; the free plan may show fewer sources.
- Google AI Mode showed sources on 74 of our 80 questions (92.5%), a higher share than any chat engine except Perplexity. It was captured once per question through Google’s results page, a different method, so it has no score on the second number.
- ChatGPT and Gemini tied on the first number, 213 of 240 each, and were two points apart on the second (37% and 35%). Treat them as level. ChatGPT ranks ahead on that margin, and its search also works without signing in.
- Claude showed sources on the fewest answers, 192 of 240, though its runs shared 68% of sites, second only to Perplexity.
The four we did not measure come after, ordered by what their makers document, then reach. Copilot and Grok both document citations, and X’s own help page calls Grok “an early version” that may be wrong, so Copilot goes first. Meta AI and DeepSeek both document web search but not how they show sources; Meta AI ranks ahead on reach, since it is built into WhatsApp, Instagram, Facebook and Messenger. We did not measure freshness, so it decided no place here.
We left out Brave Search, Kagi, Consensus, DuckDuckGo’s AI chat and You.com. They are credible, and Google’s own AI Overview for “best AI search engines” names several of them, but we have not measured them or read their documentation for this update. They are candidates for the next one.
What this means if you want your brand in these answers#
Prefer sells software for this job, so read this section with that stake in mind. The data above makes one point for brands: there is no single AI search engine to win. The engines share the biggest sites but split the long tail, and three of the four chat engines shifted their sources between runs, so a single check tells you little.
- ChatGPT cites about three sources per answer, so each slot is contested. Start with how to appear in ChatGPT search.
- Perplexity cites about 20 and leans on Reddit and LinkedIn. See how to get cited by Perplexity.
- Google AI Overviews and AI Mode cite different pages, and YouTube leads both. See the guides for AI Overviews and AI Mode.
- Gemini is the only chat engine that cited YouTube. See how to appear in Gemini.
- Claude cited no Reddit or YouTube in our test. See how to get cited by Claude.
- Copilot citations show up in Bing Webmaster Tools, so start there. See how to appear in Copilot.
- Grok reads live X posts. See how to appear in Grok.
Prefer tracks ChatGPT, Gemini, Google AI Overviews, Google AI Mode and Perplexity on every plan; Prefer Enterprise adds every other AI model you need, such as Claude, Microsoft Copilot, Grok, Meta AI, DeepSeek and Qwen. To see where your brand stands across them, run a free AI visibility audit. If you are comparing tracking software, the AI visibility tool pricing index lists each tool’s verified prices.
People also ask
The questions buyers ask next.
Make sure OAI-SearchBot can crawl your site, then earn mentions on the pages ChatGPT reads for your topic. Prefer tracks which pages ChatGPT cites for your buyers' questions and writes the content that earns a place on them, and on Prefer Managed it does the off-page work too. ChatGPT cites few sources (3.05 per answer in our API test), so each slot is contested, and two runs of the same question shared only 37% of their sites, so judge it over many runs, not one.
The ChatGPT guide →Be one of the pages it finds for the question; Prefer shows which pages Perplexity cites for your buyers' questions. Perplexity searched on every answer in our API test and cited about 20 sources each, with Reddit in 98 of 240 answers and LinkedIn in 73. Clear, crawlable pages plus a real presence on the community sites it reads give you the best chance.
The Perplexity guide →Start with ordinary Google eligibility: Google says a page must be indexed and eligible to show with a snippet, with no extra requirements. Prefer tracks AI Overviews and AI Mode separately, because the two surfaces cite different pages. In our capture only 32% of AI Mode's sources also appeared in the AI Overview for the same question, and YouTube was the most-cited site on both.
The AI Overviews guide →Both: in Prefer's API test, with search switched on, ChatGPT searched on 216 of 240 answers, and the 24 that did not search cited nothing. It searches when a question needs current information and answers from memory otherwise.
The full answer →Not yet. Google has become an AI search engine itself: it says AI Mode passed one billion monthly users in May 2026, and an AI Overview appeared on 75 of 80 questions in Prefer's September 2026 test. AI answers now sit on top of search rather than replacing it.
See the data →Questions
Asked plainly.
What is the best AI search engine in 2026?
Perplexity, for research you need to check: in Prefer's September 2026 API test it showed sources on all 240 of its answers and gave almost the same sources when asked twice. Those figures come from its sonar-pro API model, and Perplexity lists more citations per answer as a Pro benefit, so Pro is the plan for research. Google AI Mode ranks second and is already inside Google Search. ChatGPT is the best everyday assistant.
What is the best free AI search engine?
All nine engines on this list are free to use, per each maker's own pages read on 19 September 2026. Prefer's measurements came from APIs, not free apps, so we cannot rank the free apps from data. From the makers' own pages: ChatGPT's web search works without an account, Google AI Mode is free in Google Search, and Claude's free accounts can search. Paid plans add usage and extras, such as more citations per answer on Perplexity Pro.
Which AI search engine cites the most sources?
Perplexity, by a wide margin. In Prefer's September 2026 API test it cited 19.5 sources per answer and never fewer than 11. Gemini cited 8.6, Claude 4.6 and ChatGPT 3.05. On Google, AI Overviews cited 8.6 sources each and AI Mode 5.8. More sources means more to check, not a better answer.
Do AI search engines cite the same sources twice?
Only Perplexity comes close. In Prefer's September 2026 test, each engine got the same 80 questions three times, seconds apart, through its API. Two runs of the same question shared, on average, 96% of the sites cited across both on Perplexity, 68% on Claude, 37% on ChatGPT and 35% on Gemini. Gemini cited identical sites on only 2 of 200 pairs of runs.
Is ChatGPT a search engine?
It is an assistant that searches: with search switched on in Prefer's API test, it searched on 216 of 240 answers. OpenAI's help center says ChatGPT 'may search the web automatically when your question would benefit from current information', and it links sources when it does. You can also set it as Chrome's default search engine with OpenAI's extension.
Do all AI search engines read the same sources?
Mostly not, except for the biggest sites. Of 1,329 sites the four chat engines cited in Prefer's September 2026 test, 966 (72.7%) were cited by only one engine and 29 by all four, though 23 of the 25 most-cited sites were cited by three or four. Gemini cited YouTube in 113 of 240 answers; ChatGPT, Perplexity and Claude cited it in none. Our 80 questions were about AI search and marketing, and the sources move with the topic.
References
Sources, checked on the date shown.
- Prefer AI answer study, September 2026 (measured 13 September 2026) tryprefer.com
- OpenAI Help Center, Searching the web with ChatGPT (read 19 September 2026) help.openai.com
- Perplexity Help Center, What is Perplexity Pro? (read 19 September 2026) perplexity.ai
- Google Search Help, AI Mode in Google Search (read 19 September 2026) support.google.com
- Google, A new era for AI Search (19 May 2026) blog.google
- Google Search Central, AI features and your website (updated 10 December 2025, read 19 September 2026) developers.google.com
- Google, Gemini plans (read 19 September 2026) gemini.google
- Google AI for Developers, Grounding with Google Search (updated 17 September 2026) ai.google.dev
- Microsoft, Copilot in Microsoft 365 plans for individuals (read 19 September 2026) microsoft.com
- Bing Webmaster Blog, Introducing AI Performance in Bing Webmaster Tools (10 February 2026) blogs.bing.com
- Microsoft Learn, Search public data with Bing for generative answers, Copilot Studio agents (read 5 September 2026) learn.microsoft.com
- Claude Help Center, Enable and use web search (read 19 September 2026) support.claude.com
- xAI, Grok (read 19 September 2026) x.ai
- X Help Center, About Grok (read 19 September 2026) help.x.com
- xAI docs, X Search tool (read 19 September 2026) docs.x.ai
- Meta, Introducing the Meta AI app (29 April 2025) about.fb.com
- Meta, New AI tools to help you make things happen on Facebook (15 June 2026) about.fb.com
- Meta, Introducing Meta One (15 September 2026) about.fb.com
- DeepSeek-V4-Pro model card, Hugging Face (read 19 September 2026) huggingface.co
- DeepSeek Privacy Policy (last updated 10 February 2026) cdn.deepseek.com
How to choose
Answer a few questions and the shortlist writes itself.
01
Do you need to check every source?
Use Perplexity, on Pro for the fullest source lists. In Prefer's API test it showed sources on every answer and gave nearly the same sources when asked twice (96% of sites shared). If you want a shorter list that still holds steady, Claude cited 4.6 per answer and shared 68% between runs.
02
Where do you already search and work?
Use the engine built into what you already use, which is why Prefer's ranking breaks ties on reach. On Google, use AI Mode. In Microsoft 365, Copilot is already there. In WhatsApp, Instagram or Facebook, Meta AI is built in. On X, Grok reads live posts. The engine in front of you is often the one you will actually use.
03
Do you want video and community answers?
In Prefer's test, Gemini cited YouTube in 113 of 240 answers and Reddit in 93. Perplexity cited Reddit in 98 and LinkedIn in 73. Claude cited neither Reddit nor YouTube once. Pick the engine whose sources you trust for the topic.
04
Do you need to run the model yourself?
DeepSeek publishes its model weights under the MIT license, so a company can run them on its own servers, though V4-Pro has 1.6 trillion parameters. Prefer did not measure DeepSeek in its test; it ranks here for that license. Its app stores personal data in China, per its privacy policy. Everything else on this list is a hosted product.
Keep going
The 9 Best AI Brand Monitoring Tools in 2026, Compared→Prefer is the best AI brand monitoring tool: see what ChatGPT, Gemini, Perplexity and Google AI say about you and fix it, from $24/mo. 9 tools compared.Best AI Search Monitoring Tools in 2026, Compared→Prefer is the best AI search monitoring tool: prompt monitoring across ChatGPT, Gemini, Perplexity and Google's AI, from $24/mo. 7 tools compared.The 9 Best AI Search Optimization Tools in 2026, Ranked→Prefer is the best AI search optimization tool: it rewrites and drafts pages AI engines cite, then re-measures, from $24/mo. 9 tools ranked by what they do.7 Best AirOps Alternatives for AI Search Visibility (2026)→Prefer is the best AirOps alternative: five AI engines tracked and 15 AI articles from $24 a month. AirOps publishes no prices.The September 2026 AI answer study→The data behind this list: 960 answers, the method and a download.