Answer
Should you start a new AI session or continue the old one?
Continue while the answers are still specific. Start again as soon as they turn general, and bring the written state with you.
Continue while the answers stay specific to your material. Start a new one as soon as they turn general, carrying the written state forward. The signal is the quality of the specificity, not any counter the tool shows you.
Sessions have a useful life and it ends before any limit is reached. The observable signs are consistent. Answers stop referring to the specifics of your material and start describing the general case. An instruction that was being followed exactly is now followed loosely. The system re-asks something already settled, or asserts something contradicted earlier in the same session. Each of these is a symptom of the same underlying condition, and none of them appears as an error.
The cheap test is to ask something whose answer is unambiguously present in what you supplied. If it comes back correct and specific, the session is fine. If it comes back hedged, general, or drawn from common knowledge rather than from your document, the session has stopped being useful and further work will be confidently generic. This takes seconds and is worth doing before any consequential request in a long session.
Restarting is only cheap if the state was written down, which is the argument for writing as you go rather than a separate discipline. If the durable outcome of the work sits in a file, a restart costs a read. If it sits only in the conversation, restarting means losing it, and that cost is what keeps people in sessions long past the point of usefulness — the alternative appears worse and, in that moment, is.
There is a second reason to restart that has nothing to do with degradation: a change of task. When the work moves from investigating to implementing, or from drafting to reviewing, the accumulated material from the first phase is mostly irrelevant to the second and is competing with what matters. A fresh session with the conclusions of the first is sharper than a continuation, even when the continuation would have worked.
Against that, there is real value in continuity, and the page should say so. A session that has established the shape of a problem carries a working understanding that a handover only partially transfers. Restarting during genuinely exploratory work, before anything has been concluded, throws away the exploration. The rule of thumb that survives contact with practice is: continue while exploring, restart when moving from one phase to the next, and restart immediately when specificity drops.
The moment to start again is when the answers stop mentioning your particulars, which happens well before anything warns you.
Siddharth Sharma, Context Theory
Related questions
Does starting a new session lose the reasoning?
It loses the deliberation and keeps whatever was written. This is why the decisions-and-reasons habit matters: the conclusions are recoverable, the arguments are not, and the arguments are only worth preserving where a later reader might reopen the question. Recording why is what makes a restart cost a read rather than a reconstruction.
Is there a time or length at which you should always restart?
No usable one, because it depends on how much material entered the session rather than on how long it ran. A two-hour session that read three short files may be perfectly sharp, and a twenty-minute one that ingested a large document may not be. The specificity test measures the thing that matters, and any duration rule is a proxy for it.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
| Visibility lift in AI-generated answers from GEO methods | up to 40% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
Aggarwal et al., "GEO: Generative Engine Optimization", Princeton / Georgia Tech / IIT Delhi / Allen Institute for AI — KDD 2024 · GEO-bench · 10,000 queries across 8 domains · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Response | The end of a session's useful life is signalled by answers turning general, loose adherence to a previously followed instruction, and re-asking or contradicting settled points, none of which is reported as an error by any layer. | Recording these three symptoms against session length across a set of long working sessions. |
| Workflow | Asking a question whose answer is unambiguously present in the supplied material tests the property that matters, because a hedged or general response indicates the material is no longer being used rather than that the question was hard. | Running the same present-fact question early and late in a long session. |
| Buying behaviour | The cost of restarting is determined entirely by whether state was written down, which is why sessions are continued past usefulness — with nothing written, restarting genuinely does destroy the work. | Comparing restart cost for a session with a current state file against one without. |
| Software | A change of phase — investigation to implementation, drafting to review — makes a fresh session sharper than a continuation even when the continuation would work, because the first phase's accumulated material competes with the second phase's requirements. | Comparing output quality on a second-phase task performed in continuation and in a fresh session seeded with the first phase's conclusions. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one