The AI-Search Glossary: What AEO, GEO, Citations, and Visibility Actually Mean
A practitioner's glossary of AEO, GEO, prompt tracking, AI citations, and brand visibility—and the sales-deck spin that blurs them.
September 24, 2026
•
min read
Most agencies won’t tell you this: a rising CPC is not always a story about rising competition. Sometimes you are simply bidding the same amount for searches that are worth very different amounts to your business.
I used to tell clients the opposite. When I managed a home services account spending around $20k a month, I blamed competition every time average CPC crept from $4.10 to $6.30. I was wrong. Competition was part of it. The bigger problem was us. We bid $6.50 for a Tuesday morning search from a homeowner three miles away, and $6.50 for a Sunday night search from someone outside the service area with little apparent intent to buy. Google ran both auctions. Our average CPC told us nothing about which click deserved the money.
Yes, clicks cost more than they used to. WordStream data covered in Search Engine Land puts the average Google Ads CPC at $5.26 in 2025, up from $4.66 the year before, with CPCs up in 87% of industries analyzed. Legal averages $8.58 a click. Finance and home improvement sit above $7. If you run search in any of those categories, you felt it before you read it.
But an industry average cannot tell you why your account paid more. Google runs a fresh auction for every search. The price depends on the competition in that auction and the strength of the ads competing in it. A $6.50 maximum CPC does not mean every click costs $6.50. It does mean you have given the same ceiling to auctions with very different prospects of producing a customer.
That distinction matters more than a chart showing CPC up and to the right. Some expensive clicks are worth buying. Some cheap ones are not. If you cut every bid to make the average look better, you can lose the searches that were paying for the whole account. The job is to stop paying full price for weak auctions, not to make every click cheap.
Here is the cause before the effect. A human sets one keyword bid and hopes it fits a thousand searches. Machine learning estimates the value of this auction, for this searcher, in this context, then bids accordingly. Google calls this auction-time bidding. When its model expects a click to convert poorly, it can bid less or sit out. When the signals point to a stronger prospect, it can hold the bid or raise it.
If the lower bids land on clicks you did not need, average CPC can fall without sacrificing the volume that matters. That is the useful version of a CPC reduction. A cheaper account that also lost its best prospects is just a smaller account.
Think back to that $20k-a-month emergency plumbing budget. A Monday 8am search from someone in the service area can justify a different bid from a Saturday 2am search 40 miles outside it, even when both searches match the same keyword. A manual keyword bid treats those situations alike until someone reviews the reports. Auction-time bidding can treat them differently as the auctions happen. You lower CPC by paying full price less often, not by bidding less for everyone.
Google documents auction-time bidding as using signals such as device, location, time of day, search query and audience. A person cannot weigh all of that before each auction. You can, however, use the same categories to diagnose what the model is being asked to buy:
I learned to check location waste early because it showed up quickly in accounts I audited. That is not an argument for a blanket location bid adjustment on every account. It is an argument for looking at what you paid for before blaming the auction.
One more signal sits alongside bidding: relevance. Expected clickthrough rate is part of Quality Score, along with ad relevance and landing page experience. Weak relevance can make it harder to compete efficiently for the same position. No bidding model can make an irrelevant ad and landing page a perfect match for the search.
I watched a SaaS account cut brand CPC by 31% in two weeks without changing bids. We split one bloated ad group into three intent-matched groups and rewrote the landing page headline to match the query more closely. The account competed with a more relevant ad and page; the lower CPC followed. When relevance is the problem, turning the bid dial is not the first fix.
Manual adjustments and rules work in batches. You review the last seven days, set +20% for mobile and -30% for 1am to 5am, then come back next Tuesday. Auction-time bidding evaluates each search as it arrives.
A batch rule that says mobile converts 18% worse can also underbid the mobile searcher who is ready to buy now. That person may be in your service area, using a high-intent query and returning after an earlier visit. A real-time model can consider those signals together. Batch bidding averages the past. Auction-time bidding prices the present.
That does not make the model a strategist. It makes it faster at the repetitive pricing decision. You still have to decide what a conversion is worth, which searches you refuse to buy and whether the resulting customers justify the spend.
Manual bidding loses money in daylight. I managed accounts on a Monday-and-Thursday rhythm for years, and that schedule decided when problems got attention. If competition changed on Tuesday, I did not see the effect until Thursday. If weekend mobile traffic converted at half the rate, we paid through the weekend before I could act. This is what manual management costs beyond the fee: auctions keep running while the person managing them is doing something else.
Rules-based bidding looks like the fix and can repeat the mistake faster. I wrote a version of this rule a dozen times: if CPA tops $80 for three days, cut bids 20%. It sounds disciplined. But it cuts the keyword bid across auctions, including the Tuesday 10am search that was converting at $41. Volume can drop, leaving less conversion data to work with. Then the next review arrives and the rule may fire again.
I used to tell clients that rules added control. I was wrong. They add a consistent response to the condition you wrote down, even when that condition hides the difference between a good auction and a bad one. If bids change weekly and auctions change every second, you pay for the gap.
If CPC is too high, I do not start by lowering every bid. I start by finding out whether the account knows which clicks are worth buying. These four checks work in order: later decisions depend on the quality of the earlier ones.
Clean up the conversion signal. Smart Bidding learns from the conversions you mark as important. If those actions do not represent business value, it can become very efficient at buying the wrong clicks. Clean tracking, useful conversion values and enough data to learn from are the starting point, not a tidy-up task for later (how automated bidding works).
I have seen accounts with three duplicate purchase actions, two imported GA4 events and an offline import all marked primary. The model was being told that one customer action was worth far more than it was. I would consolidate the primary actions by funnel stage, check for duplication and set values with margin in mind before judging the bidding. Otherwise a CPC report can look precise while the account is buying against bad instructions. Bad tracking buys expensive clicks with confidence.
Exclude searches you never wanted. Pull 30 days of search term and location reports. Find spend that produced no conversions, then look at the actual searches and places before deciding what to block. Say $3,400 of a $20k monthly budget went to out-of-area, job-seeker and DIY queries. The central problem is not the price of those clicks. It is that you bought them.
Add appropriate negatives. If you serve specific zip codes, separate location intent where it helps you see and control the spend. Then let auction-time bidding handle the differences among auctions you do want to enter. Exclusions draw the boundary; per-auction bidding prices what remains. Do not ask a model to rescue traffic you already know has no place in the account.
Separate intent and improve relevance. A campaign mixing brand, competitor and broad prospecting makes performance harder to read. A $4 click and a $40 click may serve very different jobs. Give those intents room to be measured and managed separately. Match ad copy and landing page headlines to the searches they are meant to answer, and keep ad groups focused enough that relevance does not disappear into a pile of loosely related terms.
I check auction insights and impression share lost to rank before I change a bid. If CPC is rising while rank loss is also climbing, I want to inspect relevance and structure rather than assume the answer is a higher ceiling. Better alignment can improve how the ad competes; paying more to compensate for poor alignment just hides the work. Fix the match between search, ad and page before you buy your way around it.
Give the model a target it can work with. A Target CPA or Target ROAS set far below recent performance can squeeze volume before the account has learned how to meet it. A lower CPC might look encouraging for a week while pipeline gets worse. That is not progress.
My starting point is recent clean data, not the number someone wishes the account had delivered last quarter. I would keep an initial target near the last 30 days of performance, rather than demand an immediate 40% improvement. I also want enough conversions to learn from; around 30 in 30 days per campaign is a useful check before leaning hard on automation, not a magic switch. For a real promotion, use a seasonality adjustment instead of repeatedly moving the target mid-flight. Targets steer the machine. They do not punish it into finding better customers.
The difference should show up on a Wednesday you never see. Imagine competition crowds your high-intent searches in the morning while a prospecting campaign slips to $96 CPA and an exact-match campaign holds at $58. An autonomous system like groas can keep working on the account: respond to search-term waste, adjust where budget goes and let auction-time pricing account for differences between individual searches. No meeting has to be scheduled before someone looks at what changed.
I used to do that kind of work in a Thursday block with cold coffee and a spreadsheet. There is still a human job here: decide what the business is trying to sell, what counts as a valuable customer and which tradeoffs are acceptable. But a periodic human review is a poor way to handle work that keeps arriving between reviews. Agencies charging for that delay should have to explain it.
This will not work for everyone. If you have broken tracking, very little conversion data or an offer that changes every two weeks like the SaaS startup I mentioned, automation can learn noise and bid accordingly. Fix the signal and get enough volume to judge performance first. If you have clean data and a steady offer, pull the last 30 days and find the searches you overpaid for. Your CPC is high when good and bad searches get treated alike. Stop treating them alike.
