How to Conduct a SWOT Analysis Using ChatGPT in 15 Minutes

Why AI SWOT analyses come out generic

Ask a vague question, get an average answer. When you type “do a SWOT analysis for a mid-size logistics company,” the model has nothing specific to reason from, so it reaches for the statistical middle of every logistics article it has ever seen. That output is not wrong, exactly. It’s just true of everyone, which makes it useless to you.

This is the single biggest reason a SWOT analysis using ChatGPT gets dismissed in the room. A manager reads three bullets, recognises the register instantly, and mentally files the whole document as filler — which then taints the two or three findings in it that were genuinely sharp.

The second failure is subtler and does more damage in a real meeting. Models routinely file external market conditions under Weaknesses and internal capability gaps under Threats. Once that line blurs, the matrix stops being a decision tool — you can’t act on a threat the same way you act on a weakness, and a plan built on a mixed-up matrix targets the wrong things.

A quick way to test any SWOT: cover the labels and read the bullets. If you can’t tell which two boxes describe things you control, the analysis has already failed. That boundary is what the SBA’s competitive-analysis fundamentals lean on too — assessing your own strengths and weaknesses separately from barriers to entry and competitor pressure.

The Four-Box Discipline, explained

The Four-Box Discipline is the approach we teach at PromptPeakAI for this task: a SWOT only earns its place in a meeting when three rules hold. Load real evidence before you ask. Hold the internal and external line. Cross-match every box into an action.

Most professionals get rule one roughly right and then stop. Rules two and three are where the value is, and they’re the two that no single prompt will enforce on its own — which is exactly why a SWOT analysis using ChatGPT runs as three prompts instead of one.

Rule three is the one people find surprising. A SWOT that ends at four lists has told you what is true, which feels like progress but commits nobody to anything. Cross-matching forces each observation to meet another one and produce a decision, and the discipline is what stops the exercise becoming an annual ritual everyone quietly ignores.

What to gather before you prompt

Fifteen minutes assumes you spend the first five offline. The model can only reason about what you hand it, so the quality of your matrix is decided before you open a chat window. You don’t need a data room — you need roughly a page of specifics.

Numbers beat adjectives every time here. “Retention is strong” gives the model nothing; “we renewed 31 of 34 accounts, and the three losses were all under twelve months old” gives it a pattern to work with. The same applies to threats: name the competitor and what they beat you on, rather than writing “increased competitive pressure.”

Anonymise as you go. Swap client names for “Client A,” round revenue to a band, and strip anything a colleague would not want repeated. You lose nothing analytically — the model reasons about the shape of the number, not the number itself — and it makes the privacy question much simpler later. Our guide to giving AI context that makes output specific covers the layering in more depth.

Run your SWOT analysis using ChatGPT

Three prompts, in order, in the same conversation. Order matters: each one narrows what the next is allowed to do. Run them in a single thread so the model keeps the evidence in view.

The instruction to hold back is the important part. Left to its own devices the model will start analysing on contact, and once it has committed to a framing it tends to defend it rather than revise it.

You are helping me prepare a SWOT analysis for an internal leadership review. Do not analyse anything yet. Read the material below, then reply only with a numbered list of the factual claims you can support from it, and a short list of anything important that appears to be missing.

ORGANISATION: [what we do, in two sentences]
SIZE AND MARKET: [headcount, revenue band, geography, customer type]
PERIOD UNDER REVIEW: [e.g. the last four quarters]

WHAT IS WORKING: [figures, win rates, retention, named wins — anonymised]
WHAT IS NOT: [lost deals and the stated reasons, delays, gaps]
OUR MARKET RIGHT NOW: [named shifts, dated, with where you heard them]
WHO WE LOSE TO: [competitors and what they beat us on]

Rules: use only what I have given you. Do not add industry generalisations. Flag anything you are inferring rather than reading.

Two passes in one prompt, deliberately sequenced. Asking for all four quadrants simultaneously is what produces the boundary errors; asking for the internal pair first forces the model to settle what you control before it considers what you don’t.

Now build the matrix from the inventory above, in two separate passes.

PASS ONE — internal only. List Strengths and Weaknesses. A factor qualifies only if we could change it ourselves within twelve months. Anything driven by competitors, customers, regulation or the wider economy is excluded here.

PASS TWO — external only. List Opportunities and Threats. A factor qualifies only if it exists whether or not we act. Anything about our own capability, cost base or staffing is excluded here.

For every item give me: the factor in under twelve words, then the specific evidence from my material that supports it. Maximum five items per quadrant. If a quadrant has fewer than three well-evidenced items, say so rather than filling it.

Finish with a table: four columns, one per quadrant.

The “say so rather than filling it” line does a lot of quiet work. Without it you get five items per quadrant every time, because the model reads an empty box as a task it has failed. With it, a thin Opportunities column tells you something real — usually that you skipped the market research.

This is the stage that separates a SWOT worth presenting from a SWOT worth deleting, and it has its own section below.

Turning four boxes into four decisions

A finished SWOT describes a situation. It does not tell anyone what to do on Monday, and this is the stage most guides to a SWOT analysis using ChatGPT never reach. The conversion step — long known as a TOWS matrix — takes each internal factor and deliberately collides it with each external one, which produces four distinct families of action rather than four lists of observations.

Take the finished matrix above and cross-match it. For each pairing, propose two concrete initiatives. Every initiative must name the specific strength or weakness and the specific opportunity or threat it comes from.

GROW: a strength applied to an opportunity.
REPAIR: a weakness that must be fixed before an opportunity is reachable.
DEFEND: a strength used to reduce the impact of a threat.
RETREAT: a weakness exposed to a threat — contain, de-prioritise or exit.

For each initiative give me: a one-line description, the likely owner by role, a rough effort level of low, medium or high, and what would have to be true for it to work.

Then rank all eight by the ratio of impact to effort and tell me which three you would start this quarter, and why the other five wait.

A worked example makes the shape obvious. Say your matrix holds a strength (“we onboard clients in five days against an industry norm of three weeks”) and a threat (“our largest competitor has just cut prices by 15%”). Cross-matched, that becomes a DEFEND initiative: compete on speed-to-value in renewal conversations rather than matching the discount. Neither box alone suggested that. The pairing did.

That last instruction — rank them and justify the deferrals — is the difference between a document and a decision. Leadership teams rarely argue about whether a weakness exists. They argue about sequencing, and this gives them something specific to push against. From here, turning those initiatives into quarterly OKRs is a short step.

When your SWOT comes out wrong

Four failures account for almost everything that goes sideways with a SWOT analysis using ChatGPT, and each has a specific fix rather than a general “try again.”

The third one deserves a moment. Invented figures are the most dangerous output here because they read exactly like the real ones, and a fabricated market-share number in a board pack is a career-grade problem. Ask the model to quote the supporting evidence beside every claim, then check the ones that matter — our three-pass verification system and the explainer on why ChatGPT invents things both cover the mechanics.

On formatting: ask for the output as bulleted sections with bold headers rather than a markdown table if it’s going into slides. Tables copy badly across applications, and reformatting one by hand costs more time than the whole analysis. Our guide to formatting AI output so it’s immediately usable has the specifications that survive the trip.

Reformat the ranked initiative list for a leadership presentation.

Format: four sections, one per pairing. Each section has a bold header, then three to five short bullets. No tables. No markdown symbols. Each bullet under fifteen words, written as a decision rather than an observation.

Then write a separate 90-word summary I can read aloud as an opening, stating the single most important finding and the one decision I am asking the room to make.

Keeping strategy data confidential

Corporate strategy is close to the most sensitive material most professionals handle, which makes a SWOT analysis using ChatGPT a genuine data-handling decision rather than just a writing task. Before you paste anything, understand which account you’re pasting into — because that, not what you pay, determines how your input is treated.

The distinction that actually matters is consumer account versus business account, and most guides get this backwards. On the consumer tiers — Free, Go, Plus and Pro alike — your conversations are used to improve OpenAI’s models by default, and all four are switched off in exactly the same place: Settings, then Data Controls, then turn off “Improve the model for everyone.” Paying more does not change that default.

Business and Enterprise workspaces are the real line. On those, OpenAI states there is no training on your business data by default, and an administrator controls retention for the whole workspace rather than each person managing a personal toggle. If your SWOT touches genuinely confidential material, that is the account it belongs in — not a personal login with a setting you hope a colleague also found.

One more thing worth knowing if you’ve read older tutorials: several of them route this workflow through Canvas, the side-panel editing space. Per OpenAI’s own release notes, Canvas was removed from the current models and replaced by inline writing blocks, surviving only on legacy models until those retire. Editing your matrix now happens in the chat thread itself. If a guide tells you to open Canvas, it hasn’t been updated. Our privacy guide for professionals goes deeper on the account question.

What ChatGPT cannot decide for you

A SWOT analysis using ChatGPT compresses a slow task honestly. What it does not do is make the judgment calls, and being clear about that is what keeps the output defensible when someone senior pushes back.

There’s also a limit worth naming plainly: a SWOT built entirely from what you already believe will confirm what you already believe, faster. If every input came from your own team, expect the matrix to flatter you. Bringing in genuinely external evidence is what breaks that loop — structured competitor research for the opportunities and threats rows, and a proper risk assessment when the threats column starts looking serious.

Run a SWOT analysis using ChatGPT once on something low-stakes before you point it at the annual plan. You’ll find your own gaps in the evidence-gathering step, which is where nearly everyone finds them, and the second run takes about a third of the time. If the output is heading into a wider strategy session, facilitating that session with AI and writing the executive summary are the two natural next moves.