Perplexity's tendency to ignore follow-up questions stems from a fundamental challenge that affects all leading AI systems: large language models perform significantly worse when instructions arrive across multiple conversation turns rather than in a single complete prompt. A 2026 study from the International Conference on Learning Representations (ICLR) found that across six generation tasks, leading large language models lost an average of 39% performance when the same task was spread across multiple conversational turns.

The issue is not unique to Perplexity. Researchers at Microsoft Research and Salesforce evaluated more than 200,000 simulated conversations and confirmed that top open-weight and closed-weight models performed substantially worse when instructions arrived over several turns rather than all at once. The problem lies not in raw capability loss but in how models handle evolving context. As one researcher explained, models can make early assumptions, commit to them, and then struggle to recover when later turns clarify or redirect the task.

What Does "Ignored" Actually Mean in a Perplexity Conversation?

When users report that Perplexity ignores a follow-up, they may be describing one of several distinct failure patterns. Understanding which pattern occurred helps explain why the same visible symptom can arise from very different causes. Perplexity is not a single static chat model; it is a search product that crawls the web, synthesizes sources, maintains conversational context, and in Best mode selects a model based on the query. When these components disagree about what matters most, the latest answer can appear to ignore earlier context even when some of it is still present.

• Referent Loss: The subject disappears after pronouns like "that," "it," or "the second one." For example, asking about a laptop with strong battery life, then writing "and for AMD?" may cause the system to search for AMD generally rather than keeping the original laptop criteria. This typically responds to one short restatement of the noun.
• Constraint Loss: The topic remains correct, but a budget, date range, format, or exclusion criteria vanishes from the answer. Repeating only the non-negotiable constraints in one sentence usually repairs this pattern.
• Goal Anchoring: The model keeps optimizing for the original task after the user has changed objectives. This requires an explicit reset such as "the objective has changed from X to Y; do not continue X."
• Source Displacement: New web results dominate a point that an earlier answer established from a different source set. Freezing the source type, date range, or named sources for the next turn helps prevent this.

How to Repair a Drifting Perplexity Thread

• Restate the Subject Explicitly: When referent loss occurs, use the full noun instead of pronouns. Instead of "and for AMD?", write "Keeping the same laptop comparison, what about AMD processors?" This helps the retrieval system and model stay anchored to your original criteria.
• Create a Constraint Ledger: For conversations involving multiple requirements, state them together in one sentence at the start of a follow-up. This prevents the model from quietly dropping budget limits, date ranges, or other non-negotiable boundaries.
• Signal Intent Changes Clearly: When your goal shifts, explicitly say so. Tell Perplexity "The goal has changed from finding budget laptops to comparing gaming performance" rather than assuming the model will infer the shift from context alone.
• Freeze Source Scope When Needed: If earlier answers cited specific sources you want to keep, name them in your follow-up. This prevents live web retrieval from overwriting evidence you already established.
• Summarize and Branch: If a thread becomes generic, repetitive, or unrelated over many turns, summarize the current state and start a clean thread. This is faster than trying to recover a conversation that has drifted too far.

Why Even Strong Models Struggle With Evolving Intent?

A July 2026 Microsoft Research study pushed the question further by testing how well models handle changing user intent during a conversation. Researchers found that strong single-turn performance did not reliably transfer when users progressively revealed, revised, or redirected their goals. The conclusion was direct and sobering: today's large language models do not yet faithfully track and act on the user's evolving intent.

"When LLMs take a wrong turn in a conversation, they get lost and do not recover," stated Philippe Laban, Principal Researcher in Microsoft Research's AI Interaction and Learning group, and co-authors in their ICLR 2026 paper.

Philippe Laban, Principal Researcher, Microsoft Research

This maps closely to the experience users describe when a follow-up seems to be ignored. The model may still technically "remember" earlier text, but it has built the wrong working interpretation of it. A later correction competes with an internal plan that has already been established. This is why simply asking the same short follow-up again often produces another wrong answer; repetition does not repair the underlying state.

What Should Perplexity Users Know About Thread Continuity?

Perplexity's Help Center states that Pro Search maintains context from previous interactions and that previous questions are remembered so users can follow up naturally. However, the company does not publicly document a fixed threshold at which older turns are compressed, selectively retrieved, or no longer reliably influential. This means thread continuity is a convenience, not a guarantee that every earlier instruction remains equally active.

For users working on decision-heavy tasks, the safest workflow is to periodically restate the current objective and the few constraints that would make an answer unusable if forgotten. The useful way to troubleshoot a problem is not to ask whether Perplexity "has memory" in the abstract, but rather to identify which layer failed: the wording of the new prompt, retrieval, model routing, long-context recall, source selection, or intent tracking.

As AI answer engines become central to how people discover products and information, understanding these limitations helps both users and marketers adapt. The broader lesson is that multi-turn conversation remains a frontier challenge for all large language models, not a flaw unique to any single product. Recognizing this gap and working within it is the practical path forward.