Buyers are asking ChatGPT, Perplexity, and Gemini which software to use before they ever land on your pricing page. If you don't know what those assistants are saying about you, you're running a go-to-market motion with a blind spot in the exact place your next customer is standing. A SaaS AI visibility audit gives you a real answer instead of a guess: it tells you whether AI recommends your product, which competitors it names instead, and which pages it's actually pulling from when it does mention you.
If you've ever typed something like "how do I check if AI recommends my product" into a search bar, this guide is built to answer it properly. It walks through a full audit you can run this week with a spreadsheet and a handful of chat windows. By the end you'll have a prompt set built around your real buyers, a way to test chatgpt product mentions and every other engine your category shows up in, a scorecard that separates a mention from a citation, and a short list of what to fix first.
Step 1: Define the audit around real buying decisions
Don't start with a keyword list. Start with the questions a buyer actually types into an AI assistant before they've decided anything. Those questions come from the category you compete in, the people who choose your product, and the jobs they're hiring it to do.
Work through seven dimensions before you write a single prompt: the category you want to own, who makes the decision (founder, marketing lead, ops team, developer), the use case the software has to handle, the problem that sent the buyer looking, where they are in the decision (learning, shortlisting, comparing, choosing), the constraints that change the answer (team size, integrations, security, budget), and the alternatives they're weighing, including incumbents and manual processes.
Give more weight to prompts where the buyer is actually choosing between options. Category and comparison questions matter more here than purely educational ones, because they're the ones that shape a shortlist before you're even in the room.
When you're done with this step, you should be able to write down the category you're competing for, three to five buyer profiles, your core use cases, your real competitors, and the specific questions that could put you on or off a shortlist. That list is the spine of your audit product in AI search project, and everything after this step builds on it.
Common mistake: testing only branded prompts like "What is [your product]?" That tells you whether AI recognizes your name. It tells you nothing about whether AI would recommend you to someone who's never heard of you, which is the harder and more useful question.
DeepSmith's Discover Prompts feature can generate a starter list from your product, persona, and buyer-stage context, which saves you from staring at a blank spreadsheet. Treat it as a first draft to check against your own customers' language, not a finished list.
Step 2: Build your buyer-prompt set
Now write the actual prompts. Group them into families so you can tell later which type of question you're winning and which you're losing.
Category and shortlist prompts test whether you show up without anyone naming you first: "What's the best [category] software for [buyer type]?" or "What tools should a [persona] evaluate for [use case]?" Problem and use-case prompts start from the buyer's pain instead of a category label: "I need a tool to [problem], what should I consider?" Comparison prompts put you head to head with a named rival: "[Product] vs [Competitor] for a [team size] company." Alternative prompts track incumbents specifically, since a buyer who already names a tool they're replacing is a different signal than one browsing a category cold. Constraint prompts fold in the requirements your sales team hears every week, like SSO, white-label, or agency reporting. Vertical prompts narrow by industry or persona. Branded prompts round things out, but they shouldn't dominate the set.
A practical starting point is around 30 prompts: ten category prompts, fifteen comparison or alternative prompts, and five vertical or "best for" prompts. Expand toward 40 or 50 once your first run shows you where the real gaps sit. Keep the wording stable between runs, tag every prompt with its intent, buyer stage, and persona, and store the exact text rather than a topic label, because a rewritten prompt breaks your baseline.
Pro tip: keep branded and unbranded prompts in separate columns from the start. A high branded mention rate next to a low unbranded one is a completely different problem than a low score across both, and averaging them together hides that.
Step 3: Run the same prompts across every AI surface that matters
Pick the engines your buyers actually use, at minimum ChatGPT and Perplexity, then add Gemini, Claude, and Google's AI features where your category shows up there too. This is the step where you actually test chatgpt product mentions and every other engine side by side, so run every prompt fresh, in a new conversation each time, so an earlier answer doesn't leak context into the next one. If a prompt disappoints you, don't quietly edit it. Version it and keep the original.
Record the same fields every time: the date, the engine and plan, the country and language, whether you were logged in, the exact prompt text, the full answer, and every source the assistant showed. A screenshot or export is worth keeping too, because these tools update constantly and you'll want proof of what you actually saw.
Each engine behaves a little differently, and the differences matter for how you read the results. OpenAI has said that ChatGPT Search can surface citations and source links, but those citations can be incomplete or out of date, and being cited is never guaranteed. Perplexity describes its answers as sourced from the web in real time and backed by linked citations, so record both where your product lands in the answer and which exact page got cited. Gemini and Claude get the same treatment: same fields, same discipline, kept separate until you're ready to roll them up.
Google's AI features deserve their own row in your tracker. An AI Overview doesn't trigger for every search, and Google's own documentation notes that a page needs to be indexed and eligible for a normal search snippet before it can show up there, though meeting that bar doesn't guarantee anything. Note whether an AI Overview appeared at all, whether it named you, and which pages it leaned on.
Treat your first pass as a baseline snapshot, not a verdict. AI answers are probabilistic and can shift between runs on the same prompt. A workable rhythm is weekly checks on your highest-priority prompts, a full monthly run across the whole set, and a quarterly review where you retire stale prompts and add new competitor language.
Common mistake: running one pass and calling it done. A single answer tells you what happened once. Repeating the same prompt on a fixed schedule is what turns a snapshot into something you can actually act on.
Step 4: Score every answer at the response level
Build one row per prompt, per engine, per run. For each row, capture whether the product was mentioned by name, whether a page on your domain was cited as a source, which exact page and what type it was, where in the answer you landed (first recommendation, part of a shortlist, a passing mention, or absent entirely), which competitors showed up alongside you, the sentiment of how you were described, and whether the description was actually accurate.
Keep mention and citation as two separate columns, because they answer different questions. A product can be mentioned and cited, mentioned but not cited, cited without a meaningful mention, mentioned incorrectly, or beaten out entirely by a competitor's page. Each of those is a different problem with a different fix, and collapsing them into one number erases the distinction you need to act on.
When you're done, a second person should be able to read any row and understand exactly what the assistant said, whether you showed up, whether your own page backed it up, and what needs attention. The sentiment of how you were described belongs in that same row.
Common mistake: counting every mention as a citation. A mention is a name appearing in the text. A citation is a link to your own content doing the work. Conflating them makes your numbers look better than your actual source ownership.
Step 5: Calculate mention rate, citation rate, and share of voice
With your scorecard filled in, three numbers tell you most of what you need. Mention rate is the share of tested responses that name your product at all: responses mentioning you divided by total responses tested. Citation rate is the share that link to a page on your own domain as a source. Calculate both for the full set, then break them down by intent, buyer stage, engine, and by branded versus unbranded prompts separately, since a high branded score can mask a weak unbranded one.
SaaS share of voice AI is your visibility relative to the competitors buyers are actually weighing against you, not every company that exists in your category. A mention-based version divides your mentions by the total mentions across you and your named competitors; a citation-based version does the same with citations. Name which version you're using every time you report it, and keep the same competitor set and prompt list run over run, or the comparison stops meaning anything.
Sentiment is worth tracking, but treat it as a secondary signal rather than folding it into your core score. One large observational study analyzing over 100,000 brand-tracking responses across five major engines found sentiment noisier than simple mention status, and unbranded visibility split in a near-binary pattern, brands were either consistently mentioned or consistently absent, with less middle ground than you'd expect. The same study found branded recognition running far higher than unbranded discovery across the board, which lines up with the gap you'll likely see in your own numbers between "What is [my product]?" and "What's the best tool for [use case]?"
There's no universal benchmark for what a good mention rate or a healthy saas share of voice ai number looks like, because it depends on category maturity, how established your brand already is, and how many competitors are crowding the space. Your own first audit is the baseline. The comparison that matters is your product against its own prior result, and against the competitors you actually lose deals to.
DeepSmith's AEO area rolls mention rate, citation rate, share of voice, sentiment, and visibility trend into one dashboard with a per-platform breakdown and a competitor leaderboard, so you're not rebuilding this scorecard by hand every month. It organizes the audit and keeps the history; it doesn't promise a ranking or a sale.
Step 6: Diagnose the gap before you write anything
A missing mention is not automatically a reason to publish more blog posts. Before you touch your content calendar, figure out which kind of gap you're actually looking at.
No mention and no citation usually points to a discovery gap: check whether you even have a clear page addressing that category or use case. A mention without a citation means you're recognized but someone else owns the source the assistant is pulling from; find that page and ask whether yours is more specific or useful. A citation without a real recommendation means your page is present but not doing the convincing work; look at whether it actually helps someone compare and choose. When a competitor gets named and cited instead of you, compare their page directly against yours for specificity, freshness, and evidence. And if you're described inaccurately, the problem is usually unclear positioning or conflicting third-party descriptions rather than a content gap at all.
For every prompt you're losing, ask what page the assistant cited instead, whether it was a comparison, a review, or documentation, and whether that page answered the buyer's specific constraint better than anything you have. That's a more useful unit of diagnosis than the prompt alone, because it tells you exactly what to build or fix.
Resist the urge to publish a dozen near-identical pages just to occupy more prompts. It doesn't work, and Google has been explicit that generating volume without adding real value can run into its policies on scaled content abuse. Fix the specific page that's losing the specific prompt.
DeepSmith's Pages view shows which of your pages AI actually cites and which prompts drive those citations, and Competitor Citations shows who's winning the source placement instead of you and on which exact page. Content Map lays your coverage against competitors' by topic and funnel stage, and Opportunity Agents can turn any of these gaps into a content idea with the data point that justifies it attached. These are ways to move from diagnosis to a backlog faster, not a promise that writing a page guarantees a citation.
Common mistake: treating a missing citation as proof you need more content, full stop. The real cause might be a missing comparison page, a stale product page, weak third-party authority, or an AI feature that simply didn't trigger for that query. Diagnose first.
Step 7: Turn the audit into a recurring loop
An audit that runs once tells you where you stood on one day. The value comes from running it again on a schedule and watching what moves.
Prioritize your backlog by asking three questions of each gap: is the buyer close to actually choosing a product, is a named competitor currently winning that answer, and can you close the gap with a specific page rather than a vague content push. Work the prompts where all three line up first.
Keep the loop simple: find a prompt where you're absent or weak, identify what's currently winning it, build or fix the page that directly answers the buyer's question, check it for product accuracy and voice before it goes live, then rerun that exact prompt and record what changed. A weekly check on your priority prompts, a full monthly run, and a quarterly refresh of the prompt library itself is a sustainable cadence for most small teams.
Google Search Console will show AI-feature appearances folded into its general web search reporting, which is useful for confirming exposure, but it won't tell you which specific prompts mentioned you, which competitors showed up instead, or what got cited. That level of detail only comes from the answer-level audit you just built.
DeepSmith's Content Studio carries an idea from New Ideas through Planned Content to a finished, brand-grounded article the Writer produces with its own research, links, cover image, and metadata already attached, and Autowrite can run that on a schedule without anyone opening the app. It's a way to close the loop faster once you know exactly what to build, not a guarantee that publishing something will make an engine cite it.
What to do next
Run this saas ai visibility audit once this week with a 30-prompt starter set. Don't wait for a perfect prompt list or a bigger tool budget first; a rough baseline you can rerun beats a polished one you never repeat. Once you have that first pass scored, pick the two or three prompts where a competitor is clearly winning and you have real evidence to compete with, and fix those pages before you touch anything else.
If you'd rather not stitch this together by hand every month, DeepSmith tracks mention rate, citation rate, and share of voice across the engines your buyers use and turns the gaps it finds into a content backlog in the same platform, all part of the same workflow you'd use to audit product in AI search on your own. You can start a free trial and see your own real data and a first draft before you pay for anything.
