ChatGPT Generates 91% Unique Queries Every Run. Perplexity Generates 14%. The Same Content Strategy Won't Work for Both.

Profound ran the same 10,000 prompts through ChatGPT, Perplexity, and Copilot over 14 days. The study tracked query uniqueness, word-level overlap, intent classification, and how each platform converts user prompts into search sub-queries. What they found changes how you should think about content strategy for AI search.

The headline number: ChatGPT generates sub-queries that are 91% unique across repeated runs of the same prompt. Perplexity generates sub-queries that are only 14% unique. Copilot lands at 47%.

Run the same question through ChatGPT three times and you get three completely different search clusters. Run it through Perplexity three times and you get essentially the same keywords every time.

That is not a minor behavioral difference. It means the content that surfaces your brand in ChatGPT is built on different logic than the content that surfaces it in Perplexity, which is built on different logic than Copilot. A single content strategy optimized for one platform is leaving citations on the table for the other two.

The study was published April 30, 2026 by Profound's Jennifer Zou: "What AI engines actually search for and why ChatGPT never searches the same way twice."

What the study measured

Profound tracked 10,000 prompts across a 14-day window in late March through mid-April 2026. The methodology: query uniqueness analysis, semantic similarity scoring, word-level overlap measurement, and intent and question-type classification. Platforms covered: ChatGPT, Perplexity, and Copilot.

The core measurement was sub-query uniqueness. When a user submits a prompt, each AI platform breaks it into derivative search queries to retrieve sources. Those sub-queries were captured, compared across repeated runs of the same original prompt, and scored for how much they varied.

ChatGPT: the researcher

ChatGPT generated unique sub-queries 91% of the time across repeated prompt runs. Profound describes it as functioning like "a researcher," generating semantically similar but lexically different searches each time.

Practically, this means ChatGPT approaches the same question from different angles on every run. A prompt about CRM software might generate "best CRM tools for enterprise B2B" on one run, "CRM comparison pipeline management features 2026" on the next, and "mid-market CRM integrations Salesforce alternative" on a third. All three point at the same user intent. All three pull from different content.

Some specific behaviors from the Profound data:

ChatGPT preserves question structure 31-40% of the time, compared to Copilot's 93% conversion rate to keyword strings. For prompts containing brand names, single brands are preserved across all runs. Multiple brands are less stable; ChatGPT is the most likely of the three platforms to scatter or drop brand references across sub-queries.

For "best [X]" queries, the format your buyers use when evaluating vendors, ChatGPT retains that phrasing only 39% of the time. The other 61% of runs, it reformulates the query into something more specific.

The implication for content: you cannot rely on ranking for a single phrase to capture ChatGPT citations. The platform generates a different search cluster on most runs. The brands that surface consistently are those with broad topical coverage: multiple pieces of content covering adjacent angles of the same question. Our post on content clusters for AI citation covers the structural approach to building that kind of breadth.

ChatGPT also requires more monitoring runs to get a reliable baseline. A single visibility check for a given prompt is one draw from a large distribution of possible sub-queries. The Profound data confirms what we have seen across client work: you need rolling 30-day citation data, not point-in-time snapshots, to understand ChatGPT brand presence.

Perplexity: the keyword search engine

Perplexity generated unique sub-queries only 14% of the time. It retains 88% of the word overlap from the original prompt. Near 1:1 fanout ratio. Profound calls it "the closest to classic search" of the three platforms.

That description is accurate. When you submit a prompt to Perplexity, it searches for something very close to what you typed. The sub-queries look like search engine queries, not research angles. Keyword-native, specific, directly derived from the prompt.

This is why Perplexity's citation structure rewards precision. Pages that use exact terminology get found because Perplexity is literally running those terms. Pages that paraphrase or explain concepts without using the category's standard vocabulary miss the match.

Two other Perplexity behaviors from the study matter for brand strategy:

Perplexity retains "best [X]" query format 65% of the time, the highest of the three platforms. If a buyer searches "best project management software for marketing teams," Perplexity runs that phrase as written. Brands in listicle-format content that names specific products and their attributes for that use case get cited. Brands that only have narrative feature pages miss the match.

Perplexity also needs less frequent monitoring. Since sub-queries are stable across runs, a visibility measurement taken today will reflect what you would find tomorrow. This makes Perplexity the platform where a one-time citation audit gives you usable baseline data, rather than just one draw from a distribution.

Copilot: the compressor

Copilot sits between the other two at 47% unique sub-queries. The Profound data identifies it as "the compressor": it rewrites prompts into shorter, search-engine-native strings.

The most striking behavioral data: Copilot converts "how" questions into keyword phrases 93% of the time. A prompt like "How do I choose the right CRM for my sales team?" becomes "CRM selection sales team" or "best CRM sales workflow 2026" before it hits the search layer. The conversational framing disappears entirely.

Copilot also strips adjectives aggressively. Qualitative language that helps position your brand in human-readable content gets removed before the search query runs. "Leading," "best-in-class," "enterprise-grade": those terms rarely survive into Copilot's actual search string.

What does survive: brand names (single brands are preserved consistently), product categories, and attribute terms that function as search-native keywords.

Copilot is Bing-grounded. Its citation pool draws from Bing's web index rather than OpenAI's training data or Perplexity's own index. This creates a specific optimization path: Bing indexation quality matters differently than Google indexation quality for Copilot visibility. Brands that have optimized only for Google's crawler and index may have meaningful Bing indexation gaps that leave them out of Copilot's citation pool entirely.

One notable finding for multi-brand queries: Copilot is more stable than ChatGPT at preserving multiple brand names across sub-queries. Where ChatGPT scatters or drops additional brand names in reformulated queries, Copilot tends to retain them. This makes Copilot more predictable for competitive positioning queries; the brands that appear in Copilot's output for a "best X tools" prompt tend to appear consistently.

Microsoft Copilot grew 25.2x in 2026 by referral traffic, the fastest growth rate of any AI platform. That growth reflects adoption across Microsoft 365's 15 million paid enterprise seats. For B2B SaaS brands, Copilot is not a peripheral platform: it runs inside the tools your buyers use daily, including Outlook, Teams, Word, and Excel. The keyword-compression behavior means Bing-optimized, attribute-dense content is the path to appearing in those workflows.

The "best [X]" format matters most, but differently by platform

Across all three platforms, the most commercially valuable query format is "best [X]," the vendor evaluation prompt buyers use at the moment they are building a shortlist.

The Profound data shows how differently each platform handles it:

For Perplexity, a page titled "Best CRM for B2B Sales Teams" with a structured comparison table is directly aligned with how Perplexity searches. It will run that phrase and cite the page that answers it.

For Copilot, the same page needs Bing indexation and keyword density that survives query compression. The heading and body should use search-native attribute language: specific integrations, pricing tiers, use cases. Content that uses that vocabulary will survive Copilot's query compression and remain findable.

For ChatGPT, a single "best CRM" page is insufficient. You need content covering multiple angles: comparison pages, integration-specific pages, use-case pages, customer type pages. When ChatGPT reformulates a "best CRM" query into eight different sub-angles, each sub-angle needs a page that answers it.

Comparison pages are one of the highest-citation content types for AI search precisely because they answer multiple specific sub-angles at once. They work for all three platforms, but for different reasons.

What this means for GEO monitoring

The platform behavior differences change how you interpret visibility data.

ChatGPT citation data from a single measurement point is essentially noise. The 91% uniqueness rate means a single brand check is one draw from a large distribution of possible search clusters. Your brand may appear on 40% of runs or 90% of runs; a single measurement does not tell you which. You need rolling citation data across many runs to get a stable signal.

Perplexity and Copilot are more reliable from single checks. The 14% uniqueness rate on Perplexity means a citation measurement today predicts tomorrow's result accurately. Copilot at 47% is somewhere between the two.

This has practical implications for tool selection. AI visibility monitoring tools that report a single ChatGPT visibility percentage from one measurement per day are producing numbers with high variance. A brand scoring 60% in one daily check might score 30% the next day with the same underlying content. Daily single-run measurements create the illusion of volatility that is actually query randomization.

The more useful metric for ChatGPT: rolling 30-day citation rate across a high volume of prompt runs. That smooths the query-variance noise and gives you the actual probability your brand appears for a given category of user intent.

For tracking prompt performance across platforms, the platform difference also means you should not compare raw citation rates between ChatGPT and Perplexity as if they are equivalent measurements. Perplexity's stable queries mean its citation rate is a more precise signal. ChatGPT's variable queries mean its citation rate reflects topical coverage breadth, not precision recall.

Building content for all three simultaneously

The good news from the Profound study: the optimization approaches for each platform are compatible, not contradictory.

A brand building ChatGPT visibility through topical content clusters (multiple pieces covering adjacent angles) will automatically improve Perplexity and Copilot presence on the specific keywords those cluster pages target. Cluster pages are often the most keyword-precise pages a brand publishes, which is exactly what Perplexity's near-1:1 fanout rewards.

The gap tends to appear when brands optimize for one platform's logic exclusively:

Brands that optimize only for Perplexity (high keyword precision, exact terminology) will underperform on ChatGPT because they lack the topical breadth to surface across ChatGPT's varied sub-query clusters. A single well-optimized page does not cover 91% query variance.

Brands that optimize only for ChatGPT (broad topical coverage, many adjacent pages) may underperform on Copilot if those pages lack Bing indexation quality and attribute-dense keyword structure. Narrative-heavy content that covers a topic broadly but without search-native attribute language gets compressed into useless keyword strings by Copilot's query rewriter.

The practical sequencing for most B2B SaaS brands:

Start with Perplexity optimization basics: keyword precision, factual density, regular content refresh. This is the fastest path to stable, measurable citation presence because Perplexity's stable query behavior makes the signal clean.

Build out ChatGPT topical clusters around your highest-value category prompts. The cluster approach requires more content production but produces compounding returns because each piece covers a different sub-query angle that ChatGPT might generate.

Fix Bing indexation for Copilot visibility. Check Bing Webmaster Tools for crawl errors, coverage gaps, and indexation quality. Many brands discover their Bing indexation is a fraction of their Google indexation; that gap is the primary reason they are absent from Copilot's citation pool despite having good Google presence.

The platform optimization priority framework should now incorporate query behavior, not just audience size. Perplexity's 97% citation rate and stable queries make it the fastest platform to build measurable presence on. ChatGPT's scale and 87.4% share of AI referral traffic make it the highest-volume platform. Copilot's workplace embedding makes it strategically important for B2B research workflows. All three require distinct approaches.

FAQ

Why does ChatGPT generate such different queries every run?

ChatGPT is designed to approach questions as a researcher would, exploring multiple angles and formulations rather than running a fixed search string. The 91% uniqueness rate reflects this design. The model uses semantic similarity to determine what sub-queries are relevant but applies significant lexical variation across runs. This produces broad information retrieval but makes citation measurement harder than platforms that run stable queries.

Does Perplexity's keyword-search behavior make it easier to rank on?

In some ways, yes. Because Perplexity runs near-identical sub-queries on repeat runs, optimizing for specific target keywords produces predictable results. A page that matches Perplexity's query for a given topic will appear consistently. The challenge is that Perplexity's 97% citation rate also means competition for cited positions is higher; more brands are optimizing for it than most GEO practitioners realize.

How does Copilot's Bing dependency affect my existing SEO work?

Brands with strong Google SEO may still have poor Copilot presence if their Bing indexation is weak. Google and Bing have separate crawl budgets, separate indexation quality assessments, and separate algorithms. A brand ranking in Google's top 10 for a category keyword may not be indexed at all by Bing for the same term. Checking Bing Webmaster Tools is a necessary step for any B2B brand that cares about Copilot visibility in Microsoft 365 workflows.

How often should I monitor ChatGPT versus Perplexity?

ChatGPT needs high-frequency, high-volume monitoring to get a stable signal due to 91% query uniqueness. Running a single daily check produces noisy, unreliable data. A rolling 30-day window with many prompt runs per day gives you the actual distribution of your citation presence. Perplexity's stable 14% uniqueness means less frequent monitoring produces reliable data; a weekly or bi-weekly check is sufficient for most brands.

Does this mean my Perplexity ranking will stay stable once I achieve it?

More stable than ChatGPT, yes, but not permanently fixed. Citation drift still occurs on Perplexity due to content freshness decay, model updates, and competitive changes. Perplexity's stable query behavior means that when you achieve a citation, it persists across repeated runs of the same query. But fresh competitive content, Perplexity's content refresh cycle, and periodic model updates still shift which sources get cited over time.

The content strategy gap most brands have

Most B2B SaaS brands running GEO programs in 2026 are treating AI search as a single channel. They optimize for ChatGPT because it has the most traffic share. They may add Perplexity because it is visible and growing. Copilot often gets skipped because it is harder to see and harder to measure.

The Profound data reframes this. Each platform runs a fundamentally different retrieval process. Treating them as one channel means your content is optimized for one retrieval logic while operating on two or three others that require different approaches.

The brands that will build durable AI citation presence in the next 12 months are those building platform-specific content programs: topical clusters for ChatGPT's variable queries, keyword-precise pages for Perplexity's stable searches, and Bing-indexed attribute-dense content for Copilot's compressor behavior. The overlap between these approaches is real; a well-built topical cluster still helps Perplexity and Copilot. But the gaps between platform-specific optimization and generic content are significant enough to matter in competitive categories.

Peec AI's Momentum case study from April 2026 found 10x AI visibility improvement from 100 articles optimized against specific prompts across platforms. The prompt-specific optimization approach is the practical application of what the Profound study explains: different platforms require different content shapes to appear consistently in their citation pools.