Why Is My Google Traffic Declining? AI Answers Are Citing Competitors

Your Google chart rolled over, and the usual suspects are lined up waiting for the blame. The blog slowed down. A competitor bought links. The algorithm "changed." Sometimes that is the story. A lot of the time, in Phoenix and the East Valley, the click never left the results page. An AI answer spent it.

Google's AI Overviews, AI Mode, and the chat tools people open instead of a browser — ChatGPT, Gemini, Perplexity, Copilot — are now where a chunk of research happens. They do not rank you in the old sense. They quote a short list and move on. If your page is not in that list, the model will cheerfully name the competitor whose page was easier to lift. Traffic declines. The competitor did not have to outrank you. They had to be more quotable.

This is a read of the public evidence, not a promise that a new file or a posting streak will put you back on a 2022 traffic chart. The studies disagree on the size of the hit. They do not disagree on the direction.

The click got answered before the link

Pew Research Center watched real browsing, not a rank tracker. In March 2025, about 900 U.S. adults produced 68,879 Google searches. Roughly 18% of those searches showed an AI summary. On visits with a summary, people clicked a traditional result 8% of the time. On visits without one, they clicked 15% of the time. The sources printed inside the summary were clicked on about 1% of visits. People were also more likely to end the session: 26% after a summary, 16% without one.

Pew is careful, and you should be too. This is observational. The searches that earn a summary are not a random half of all searches. The longer write-up of the same panel, on arXiv, finds summaries more often when the query is longer, starts with a question word, or contains both a noun and a verb. "How often should I flush a tankless heater in Gilbert" is a different animal from "tankless heater install near me."

• No AI summary15%
• AI summary shown8%
• Clicked a source inside the summary1%

Pew Research Center, March 2025 browsing panel of U.S. adults. Observational, not a Phoenix-only sample.

Seer Interactive tracked informational queries for 42 organizations from June 2024 through September 2025: 3,119 terms, 25.1 million organic impressions. Organic click-through on queries that showed an AI Overview fell from 1.76% to 0.61%, which they report as a 61% drop. Queries with no overview also fell, from about 2.7% to 1.62%. People are clicking less even when the overview is absent. Seer also says brands cited inside an overview had higher CTR than brands that were not, and then says the quiet part out loud: that may be because stronger sites get cited, not because the citation caused the click. Do not let a salesperson turn "35% more clicks when cited" into a contract.

For a Phoenix service business the split is practical.

• Research queries are where the chart breaks. Cost, comparisons, "best," "how does this work," "is this worth it." The answer can be complete without you.
• Job queries still need a business: a name, a service area, a phone, a review count. The Map pack did not retire because a chatbot learned to explain hard water.
• Your analytics tool cannot see the answer that named someone else and produced zero sessions. The absence is the data.

If organic sessions are down and the queries that fell are questions, you are not looking at a penalty. You are looking at a product change. Google is serving the answer. The websites are the footnotes, and footnotes get fewer visits. That is the whole "traffic decline" for a lot of local sites that were never ranking for "emergency plumber Phoenix" in the first place. They were ranking for explainers. Explainers are what the summary eats.

Who gets named when the model picks a local business

Ask a chat tool who to call for a water heater in Tempe and you will not get ten blue links. You will get a few names, sometimes with a reason that sounds like it read a page. The businesses that show up tend to share a boring trait: a page that states the offer in plain sentences. Service, area, constraint, proof. The homepage that says "quality you can trust, serving the Valley" gives the model nothing to prefer you over the shop that wrote "same-week tankless replacement in Tempe and Chandler, permit pulled, old unit hauled off."

That is the theft owners feel, and it is not mystical. The model is compressing the web. Compression rewards the page that was already a clean answer. Your competitor did not steal a ranking. Their sentence was easier to repeat without lying.

A few patterns, from watching what these systems actually print, not from a secret prompt list:

• They name the unambiguous business. One name, one service, cities you can point at. Three brands and a holding company confuse the citation, so the model picks the competitor who looks like a single shop.
• They repeat specifics that appear in more than one place: the site, the Google Business Profile, a review that mentions the job. If those three disagree, you become the risky quote.
• They skip doorway pages. A stack of suburb URLs with the city swapped is not five answers. It is one answer wearing hats. Google has punished that shape for years. An LLM does not need a manual action to ignore it. The service-area page problem is the same problem in a new interface.
• They will recommend a competitor you have never "lost" to in the Map pack because the question was never a Map pack question. The research query and the "who should I hire" query are different. You can own one and be invisible on the other.

None of this requires the model to like you. It requires the page to be true, specific, and indexable. Google's own note on AI features is dull on purpose: to be eligible as a supporting link in AI Overviews or AI Mode, the page has to be indexed and eligible for a snippet. No extra file. No special schema type. If the important words are trapped in an image or a tab the crawler never renders, you are not in the pile the model is choosing from.

What LLMs keep quoting

Forget the folklore about "writing for robots." The pages that get lifted look like pages a busy owner would actually read.

A direct first sentence. The heading asks the question. The next line answers it. Then you can be clever. Models extract. They do not scroll for your brand story. If the answer is in paragraph six, you donated the citation to whoever put it in paragraph one.

Nouns you can check. Cities, the service, the hours, what you refuse to do. "East Valley" is a region. "We dispatch to Mesa, Gilbert, and Chandler, and we do not run same-day west of the I-17" is a fact. Facts get repeated. Slogans get dropped.

One identity. The business name on the site, on the profile, and in the footer should be the same string. Nicknames are how you get cited as two half-businesses and then as neither. NAP consistency was a local SEO chore. It is now how a model decides you are a real entity and not a keyword page.

Proof that is on the page. Reviews you can quote, photos of the work, a process in order. Invented ROI is not just a trust problem with humans. It is a sentence the model might repeat, and then you own the lie. We do not publish client rank claims for that reason. You should not hand a model one either.

Sources when you use a number. Pew, Seer, Google's docs. A local page that says "most homeowners choose us" with no basis is commodity text. Google's generative AI guide tells site owners to aim at unique, non-commodity content. A rewritten version of a national "why is my traffic down" post, with Phoenix sprinkled on top, is commodity. A page that explains what your trade actually does in this climate is not.

Dates that mean something. "Updated" is useful when the offer, the law, or the season changed. Bumping the date on a stale price is how you teach both Google and a chatbot that your timestamps are costume jewelry.

Visible text that matches the schema. Google has said structured data should match what people can see, and that there is no special AI schema to add. Markup that describes a service you do not show on the page is a magic trick. Models and quality raters are both bad audiences for magic tricks.

How often you actually need to post

There is no study that says "publish four times a week and ChatGPT will cite you." Anyone selling that cadence invented it. What the systems show is narrower. Question-shaped queries get summaries. Stale, vague, or duplicated pages do not become the sentence in the summary just because you shipped another one yesterday.

So the cadence is about truth and coverage, not a streak.

For a Phoenix or East Valley service company, the pages that deserve a calendar are the ones that match how people ask:

1. The service pages that name the job, the area, and the constraint. Review them when the work changes. A quarterly read is plenty if the offer is stable. Rewrite them the week the offer is not.
2. The Google Business Profile, which is still the object on the Map pack. Hours before a holiday weekend. Categories that match the pages. Reviews you actually earned. A blog post does not fix a profile that says you are closed.
3. A short list of answer pages for the questions your calls already contain. Price bands you are willing to say out loud. What the job includes. What it does not. How scheduling works in summer versus January. Those pages can sit for months if they stay accurate.
4. New posts when the question is new. A change in equipment, a city you started serving, a mistake you keep seeing on jobs. That is a reason to publish. "We have not posted this week" is not.

Three solid posts a week is a publishing-company pace. It is what a site like this one can do when each post has a different job. It is a ridiculous pace for a plumbing company whose last four posts were the same checklist with the city name changed. That pattern is how you train Google to treat you as a doorway site and how you give an LLM five URLs that all say nothing. Fewer pages, each one quotable, beat a calendar full of swaps.

If you want a rhythm an owner can keep: one pass a quarter on the money pages, and a new answer only when the phone has asked for it twice. Refresh the sentence that is now wrong. Leave the sentence that is still right. Models are not impressed by your editorial calendar. They are impressed by a page that is still correct on Thursday.

llms.txt is a map, not a ranking trick

You will see the file sold as LLM.txt, llms.txt, or "the new robots.txt." The proposal that exists is /llms.txt, from Answer.AI in September 2024, documented at llmstxt.org. The shape is simple. An H1 with the site or company name. A blockquote that says what the place actually does. Optional notes. Then H2 sections whose lists are real links, each with a name and a short note about why the link exists.

It is for agents that go looking for a curated overview, mostly so they do not have to guess which of your 200 URLs matters. The proposal's own authors expected it to matter more for documentation than for a local plumber, and more at answer time than in a training crawl.

Google's position is the one that should kill the fantasy. In the generative AI guide, and in the June 15, 2026 update to that guide, Google says Search does not use these files. Creating one will not help or hurt your rankings or your visibility in Google's AI features. Google may crawl the file the way it crawls other non-HTML files. Crawling is not a special treatment. Search Central's AI features page says the same thing in shorter form: no new machine-readable file is required.

What about ChatGPT, Perplexity, Claude, Gemini? There is no matching public commitment that those products fetch /llms.txt on a consumer "who should I hire" question. Correlation write-ups that went looking for a citation bump have not found one that survives contact with "the site was strong anyway." Treat any guru screenshot that says the file "forces" a mention as advertising.

So why have one at all?

• It is a small, honest index of the pages you want a careful agent to open: the service, the contact, the definitions you actually use.
• It keeps you from accidentally featuring the city-swap URLs as if they were the good pages.
• It costs little if you generate it from the same list a human uses in the footer. It costs a lot if you maintain a second, prettier version of the site that drifts from the real one.

We publish one at arizonastrategy.com/llms.txt. It names the services, the East Valley market, and the pages worth citing. It is not on the site because Google asked. It is on the site so a tool that does read the convention finds the canonical offer instead of a random blog URL. If you copy the idea, copy the discipline. Every link in the file should be a page you would hand a customer. If you would be embarrassed to have the model quote it, do not list it.

A useful file, in practice:

• H1 is the business name, spelled the way the profile spells it.
• The blockquote says who you serve and where. One or two sentences. Not a keyword block.
• Links go to the service pages, the contact path, and a few answers that are still true. Notes say what the page is for.
• You delete links to outdated offers the same week you kill the offer.

That is the whole trick. There is no schema ritual hiding under it.

What still sends the call

Traffic was never the job. The job is a lead from someone in Phoenix, Scottsdale, Tempe, Mesa, Chandler, or Gilbert who needs the thing you do. AI answers change which queries can produce that lead without a visit. They do not change the rest of the machine.

Keep the Map pack honest. Categories, services, hours, and photos that match the site. The profile checklist is still the right punch list for the queries an overview does not eat. Read how near-me queries behave before you decide the pack is "over" because a blog chart dipped.

Keep the service pages faster than the explanation. If someone does click out of an overview, or Googles your name after a chat tool mentioned you, the page has a few seconds. Offer, proof, a way to call. A slow page full of the same paragraph the model already showed them is how you lose the visit you were lucky to get.

Keep a line on paid search for the jobs you can staff. Clicks are down there too, on Seer's numbers, including on queries with no overview. That is not a reason to shut ads off. It is a reason to know which queries still turn into calls before you raise a budget. Organic and paid are both footnotes on some searches now. The businesses that get named in the footnote and still answer the phone are the ones taking the work.

And keep the measurement adult. Pull Search Console queries where impressions held and clicks fell. Those are your summary victims. Pull queries where both fell. Those are a different illness: indexation, a lost page, a competitor who actually took the ranking. Mixing them into one "Google traffic is dying" sentence is how you buy the wrong project.

If the decline is real and the queries are questions, the work is to become the sentence. Say what you do, where, and what you will not do. Put that on a page Google can snippet. Point an llms.txt at that page if you want the map. Then watch calls and branded search, not the session chart you fell in love with in 2021.

The shops getting named are not secretly prompting their way in. They wrote the page the model could repeat without inventing a detail. You can write that page. You cannot spreadsheet your way to it with another city in the title.