Perplexity Gets 1.3%. Stop Weighting It Like Google.
Your AI visibility tool hands you one number every month, averaged across six engines, and that number is the problem. For most local and service businesses, the correct move is to weight ruthlessly toward Google's AI layer first, ChatGPT second, and drop Perplexity to a monitor-only footnote you glance at quarterly. Every top-ranking guide on this topic says the opposite, some version of "optimize for all of them," and the tracking tools are built to keep you paying for an equal-weight average. By the end of this you will have the weighting, the sourced numbers behind it, and a one-page sheet you can hand your agency on Monday.
Nobody Will Tell You Which Engines to Skip, So Your Average Hides the Only Metric That Converts
Read the three pages currently ranking for this question and count how many tell a plumber in Sacramento which engine to ignore. Zero. They list every assistant, describe each one's quirks, and leave you with a to-do list six items long, which for an owner running a business with four staff is the same as no list at all. A local or service business that lives on Google Maps and phone calls gets the overwhelming majority of its AI-driven discovery from Google's AI Overviews plus AI Mode, then a distant second from ChatGPT, and effectively nothing from Perplexity. Averaging those together produces a score that moves for reasons that have nothing to do with your phone ringing.
I have watched search platforms rise, peak and quietly die since 2008, and I run an AI visibility mention and citation tracker I built myself, which means I can weight engines against a real customer base instead of a vendor's leaderboard. Across 300+ businesses in the USA, Canada, UK, Singapore, Australia and New Zealand, mostly medical, local services and ecommerce, the pattern in where buyers actually show up is not subtle. It is the same focus-where-the-demand-is judgment that kept a London ADHD clinic booked solid for three straight months, to the point where they hired more specialists and outsourced the overflow, because we went deep on the channel their patients were already using instead of spreading thin across every channel that existed.
Translate that into money. If your service page used to catch a click on 43% of the searches that now resolve inside an answer box, the revenue from those searches does not disappear, it moves to whoever the answer box names. Your equal-weight visibility score does not tell you whether that is you, because a strong month on Perplexity can pull the average up while the engine that feeds your calendar quietly drops you.
The Equal-Weight Habit Came From the Software, Not From Your Customers
Tool vendors have a reasonable commercial motive for scoring every engine the same: more engines tracked reads as more value delivered, and a single blended score is easy to put on a dashboard and a renewal invoice. Owners inherit that framing without ever choosing it. Then the framing gets enforced socially, because the agency reporting to you would rather show movement on six lines than defend a decision to stop watching one.
The counter-reaction went too far the other way. Ross Hudgens, CEO of Siege Media, posted on LinkedIn in early September 2026 that everyone should strip Perplexity out entirely, and a lot of people reshared it as settled.
Deleting a data source is a different decision from down-ranking it, and the first one is hard to undo. Greg Jarboe of SEO-PR, writing in Search Engine Journal on September 9, 2026, made the distinction that most of the reshares skipped.
Jarboe backs it with a memory worth borrowing. In 2002 the standard agency report tracked Yahoo, Excite, Lycos, AltaVista and Ask Jeeves. Google was not on the list. His line about it is the cheapest insurance policy in this whole argument: "the platform you stop tracking may be the one you wish you hadn't." Keeping a monitor-only row in a spreadsheet costs you nothing. Rebuilding two years of trend data because you deleted the row costs you the ability to spot the next Google while it is still small.
The Share Numbers Point Somewhere Specific, and It Is Not Where the Noise Is
Similarweb measured share of visits across seven AI assistants in May 2026, and the distribution is lopsided enough that it should change what you look at first.
| AI assistant | Share of visits, May 2026 | What that means for your tracking |
|---|---|---|
| ChatGPT | 53.9% | Track on its own line, never blended. |
| Gemini | 27.9% | Track on its own line, and rising. |
| Claude | 9.2% | Promote it only if you sell B2B or to enterprise. |
| DeepSeek | 4.1% | Monitor only. |
| Grok | 2.4% | Monitor only. |
| Perplexity | 1.3% | Monitor only, and stop reporting it as a headline. |
| Copilot | 1.3% | Promote it if your buyers work inside Microsoft 365. |
Perplexity's referral share is falling too, not just small. StatCounter's worldwide AI chatbot referral data has it dropping from 7.91% in June 2026 to 4.31% in August 2026, while Gemini climbed from 7.94% to 10.9% over the same three months. So a tracker giving those two engines the same weight in your monthly score is telling you a story that reverses reality. If your agency's report shows AI visibility "up 4 points" and the gain sits on the falling engine, you are paying for a number that will not survive contact with your booking calendar.
| Referral share | June 2026 | August 2026 | Direction |
|---|---|---|---|
| Perplexity | 7.91% | 4.31% | Nearly halved in three months. |
| Gemini | 7.94% | 10.9% | Passed Perplexity going the other way. |
The scale difference is bigger still once you count Google's AI layer properly. AI Overviews reach about 2.5 billion users a month. AI Mode has passed 1 billion monthly users and roughly 75 million daily, and around 93% of AI Mode sessions end without a click. OpenAI reported 1 billion-plus users across its products at the end of July 2026, and the Gemini app hit 1 billion monthly in August 2026. Against those, a 1.3% assistant is a rounding error in your pipeline, which is exactly why it should not get a seventh of your attention. This is the prioritization rule that sits on top of an argument I have made before, that there is no single AI visibility score because each engine cites different sources. Different sources means different fixes, and different fixes deserve different budgets.
The mirror-image mistake is real and I do not want to trade one error for another. Checking only ChatGPT, which plenty of owners do because it is the name they know, misses the 27.9% of assistant visits that now go to Gemini. Under-tracking and equal-weight tracking are both failures of the same skill, which is deciding where your buyers are before you decide what to measure.
Rebuild Your Tracking in One Afternoon, Using Three Tiers
Jarboe's three-tier structure is the cleanest version of this I have seen published, and it maps almost exactly onto what I would have drawn on a whiteboard for a clinic or a home services company. Tier 1 is ChatGPT and Gemini, tracked separately and never averaged, with Claude joining that tier if you sell B2B or to enterprise buyers. Tier 2 is Google AI Overviews and AI Mode, treated as a separate and much larger discovery layer rather than "just another LLM," plus Copilot when your customers live inside Microsoft 365. Tier 3 is Perplexity, Grok and DeepSeek: monitored, not weighted equally, not ignored.
The work you do off the back of this weighting is mostly work you control, which is the part owners find hardest to believe. Yext analyzed 6.8 million AI citations in 2026 and found that 86% of them come from brand-managed sources: 44% from first-party sites, 42% from listings, 8% from reviews and social, and about 2% from Reddit. That study directly contradicts the widely repeated claim that 68% of citations come from third-party sources, a figure I would not put in a report now that a 6.8 million citation dataset says the opposite. Your own pages and your own listings are the lever. That is also why the engine quietly taking local leads is Google rather than ChatGPT for most service businesses, since Google's AI answers draw on the same listing and site signals you already own.
Check the Right Things Monthly, and Retire the Numbers That Flatter You
Frequency matters as much as weighting. Owners check daily when they are anxious and never when they are busy, and both patterns produce noise. Weekly is right for one thing only: whether Google's AI answers for your three money queries still mention you. Everything else is a monthly or quarterly rhythm, because AI answers vary run to run and a single bad sample is not a trend.
Three numbers deserve to be retired from your report. The blended AI visibility score, for every reason above. Any mention-rate figure that does not state how many prompts were run, because one lucky answer is a coin flip and asking an assistant once tells you nothing. And referral traffic from AI assistants read as a measure of influence, since around 93% of AI Mode sessions end without a click, meaning the customer can read about you, decide, and call you without ever appearing in your analytics as an AI visit.
Weighting is not a permanent verdict. It is a quarterly decision you revisit with the share data in front of you, which is exactly why the monitor-only block stays alive. Gemini went from 7.94% to 10.9% of referral share in three months. Something else can do the same.
Frequently Asked Questions
Should I remove Perplexity from my AI tracking tool?
No, demote it instead. Perplexity sat at 1.3% share of visits across seven AI assistants in Similarweb's May 2026 data, and its worldwide referral share fell from 7.91% in June 2026 to 4.31% in August 2026 per StatCounter, so it should not carry the same weight as ChatGPT or Gemini in your report. Keep collecting the data in a monitor-only block you read quarterly, because deleting the row also deletes the trend line you would need if it rebounded. In 2002 the standard agency report tracked Yahoo, Excite, Lycos, AltaVista and Ask Jeeves, and Google was not on it.
Which AI platform matters most for a local service business?
Google's AI layer, by a wide margin. AI Overviews reach about 2.5 billion users a month and appeared in 43% of US searches by May 2026, up from 15% a year earlier, and AI Mode has passed 1 billion monthly users. ChatGPT comes second at 53.9% of assistant visits in Similarweb's May 2026 figures, with Gemini at 27.9%. If you take bookings by phone or through Google Maps, put Google's AI answers at the top of your report and treat everything else as secondary.
What should I actually do to get cited by AI answers?
Work on the assets you control, because Yext's 2026 study of 6.8 million AI citations found 86% come from brand-managed sources: 44% first-party websites, 42% listings, 8% reviews and social, and about 2% Reddit. That means your own service pages, your business listings and your review profiles carry most of the weight, not scattered third-party mentions. Fix your pages so they answer the question a buyer would type, keep every listing accurate and consistent, and keep reviews coming in. It also means the widely repeated claim that 68% of citations come from third-party sources does not hold up against that dataset.
Somebody is selling you an average. Averages exist to make uneven things look comparable, which is useful when the things really are comparable and dishonest when one of them reaches billions of people and another reaches 1.3% of assistant visits. If your last report showed a single AI visibility number and you felt reassured by it, that is worth sitting with. Book a call and we will rebuild the weighting around your actual customer base.
