Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It

Jonghyun Song; Haewon Park; Jeonghoon Shim; Woojung Song; Yohan Jo

cs.CLOct 2, 2026

Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It

Authors: Jonghyun Song, Haewon Park, Jeonghoon Shim, Woojung Song, Yohan Jo

Organizations: Graduate School of Data Science, Seoul National University

Abstract

As LLM agents decide on users' behalf which product to buy, which hotel to book, or which paper to cite, a preference for items from certain sources (the sites or services they come from) shapes what users receive and which sources are selected. We study source preference in end-to-end search with 12 agent models across three domains. Comparing items from different sources that satisfy the same requirements at the same position, we find that each model prefers some sources and avoids others in every domain, largely agreeing on which. This preference can outweigh how well items satisfy the request: an item satisfying one requirement fewer is selected about two-thirds of the time when it comes from a preferred source and the better one from a dispreferred source, but almost never in the reverse case. The information identifying an item's source affects selection by itself: hiding it weakens the preference, and relabeling an item with a preferred source raises its selection rate. We test two routes to this preference: training that rewards better items can make a source a shortcut for requirement satisfaction, and missing information can trigger preconceptions about the source. Supplying missing information or a prompt countering these preconceptions reduces source preference.

Figures & tables

Figure 1: Overview. Top: agent setting (§ 2 ). Bottom: measuring source preference (§ 3 ).

		Shopping				Accommodation				Scholar
Family	Model	Amazon	Walmart	eBay	Target	Tripadvisor	Facebook	Expedia	Booking.com	arXiv	ScienceDirect	OpenReview	Medium
		12%	6%	5%	3%	9%	7%	4%	3%	12%	4%	4%	4%
Qwen	3.5-27B	$+26$	$+20$	$+1$	$+10$	$+2$	$-46$	$-18$	$+21$	$+32$	$+14$	$+18$	$-24$
	3-30B-A3B	$+16$	$+21$	$+3$	$+8$	$-6$	$-35$	$-18$	$+8$	$+23$	$+7$	$+3$	$-13$
	2.5-32B	$+15$	$+17$	$+8$	$+4$	$-10$	$-32$	$-8$	$+8$	$+23$	$+8$	$+8$	$-7$
GPT	5.4-nano	$-6$	$+11$	$-18$	$+5$	$-14$	$-52$	$-5$	$+11$	$+34$	$+11$	$+29$	$-24$

Table 1: Preference score

\tau_{s}

(percentage points), with sources ordered by frequency; the row below the source names gives each source’s share of all results, pooled over models. Cells are colored by label, determined after FDR correction (§ 3 ): Source

+

red , Source

-

blue , with intensity proportional to

|\tau_{s}|

; Source

0

entries are shown in gray .

Figure 2: Mean

\tau_{s}

across models (points), with range (bars) and majority label (color).

Figure 3: Inversion rates of Eq. 2 . Bars: 95% bootstrap confidence intervals; dashed line: 50%. Each point rests on a median of about 2,000 comparisons with exactly one item selected; the missing point has no pairs to compare. Exact values in Table F.5 .

Condition	Title / content	URL
Hidden	source names masked	hidden
URL-only	source names masked	shown
Original	original text	shown

Table 2: Source-disclosure conditions. All con- ditions use the same search results

Figure 4: Source preference across disclosure conditions. Each line shows a model’s preference score

\tau

, averaged equally across the three domains and expressed in percentage points, for (a) Source

+

or (b) Source

-

. The Hidden

\rightarrow

URL-only transition restores the URL; the URL-only

\rightarrow

Original transition additionally restores source names in the title and content.

Model	Shopping	Accommodation	Scholar
GPT-5.4-nano	$+19.2^{\ddagger}$	$+24.1^{\ddagger}$	$+18.6^{\ddagger}$
GPT-5.6-Luna	$+27.0^{\ddagger}$	$+20.1^{\ddagger}$	$+14.1^{\ddagger}$
GLM-5.3-Flash	$+25.8^{\ddagger}$	$+12.7^{\ddagger}$	$+10.4^{\ddagger}$
DeepSeek-v4-Flash	$+26.5^{\ddagger}$	$+11.8^{\ddagger}$	$+9.7^{\ddagger}$
Gemini-3.7-Flash	$+38.5^{\ddagger}$	$+13.3$	$+13.5^{\ddagger}$
Llama-4-Maverick	$+64.7^{\ddagger}$	$+14.9^{\ddagger}$	$+21.4^{\ddagger}$

Table 3: Effect of displayed explicit source information. Values show

T_{+}-T_{-}

(pp): the increase in selection under Source

+

versus Source

-

when the two sources are swapped.

\dagger

p<0.01

;

\ddagger

p<0.001

against zero.

	Pre-DPO	DPO ( $n=5{,}000$ )
Model	( $n=0$ )	Balanced	Aligned	Reversed
Qwen3.5-9B	$50.6\pm 1.0$	$50.1\pm 0.5$	$70.1\pm 0.5$	$28.5\pm 1.4$
Llama-3.1-8B	$49.8\pm 0.4$	$49.4\pm 0.1$	$73.8\pm 0.4$	$26.5\pm 0.5$
Tulu-3-8B-SFT	$51.0\pm 2.3$	$50.1\pm 0.4$	$75.1\pm 0.6$	$24.5\pm 1.0$

Table 4: Target-source selection rate (%) on equal-satisfaction test pairs, averaged across three fake source pairs. Entries report the mean

\pm

sample standard deviation across fake source pairs. Balanced, Aligned, and Reversed pair the target source with the preferred response in 50%, 80%, and 20% of training pairs, respectively.

	Base (%)	Change from Base (pp)
Model		Same Price	General	Designate
GPT-5.4-nano	68.8	$-3.9$	$-1.1$	$-17.1$
Llama-4-Maverick	95.7	$-7.1$	$+1.0$	$-12.5$
Llama-3.1-8B-It	83.9	$-17.7$	$-0.6$	$-7.4$
Qwen3.5-27B	96.4	$-2.8$	$+1.3$	$-22.5$
Tulu-3-8B	87.6	$-28.3$	$-0.1$	$-11.8$

Table 6: Source preference under added information or instructions. Base reports the Source

+

selection rate. Each intervention is applied independently; changes are measured in percentage points (pp).

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Domain	Benchmark	#Requests	Construction
Shopping	WebShop	1,500	First 1,500 from shuffled goals
Accommodation	HotelQuEST	786	All 214 + 572 location-substituted variants
Scholar	ScholarGym	2,536	All queries
Total		4,822

Appendix

Table C.1: Composition of the evaluation request sets across the three domains.

			Agreement		Kripp. $\alpha$		Gwet’s AC1
Domain	# Items	# Features	H–H	H–L	H–H	H–L	H–H	H–L
Shopping	375	1,265	.86	.88	.72	.75	.72	.76
Accommodation	375	1,198	.86	.89	.69	.74	.75	.80
Scholar	100	272	.85	.85	.67	.65	.72	.73

Appendix

Table D.1: Feature-level human and judge agreement. H–H denotes human–human agreement. H–L denotes pooled agreement between Qwen3.8 and each human annotation.

Model	Requests	Pairs	Comparisons	Rated	Source $+$	Source $-$	Source $0$
Shopping
GPT-5.4-nano	1,373	9,857	86,362	110	5	17	88
GPT-5.6-Luna	1,340	11,594	105,532	106	12	32	62
Gemini-3.7-Flash	785	7,053	50,866	77	7	8	62
GLM-5.3-Flash	1,318	10,887	82,077	138	12	21	105
Deepseek-v4-flash	1,094	8,232	61,039	95	8	8	79

Appendix

Table E.1: Data behind each model and domain: requests contributing at least one comparison, requirement-matched pairs, position comparisons after excluding runs with no selection, rated sources (at least 15 requirement-matched pairs), and how many of them are labeled Source

+

, Source

-

, and Source

0

	Shopping		Accommodation		Scholar
Model	$r$	MAR	$r$	MAR	$r$	MAR
GPT-5.4-nano	0.64	0.038	0.76	0.052	0.69	0.048
GPT-5.6-Luna	0.72	0.028	0.72	0.056	0.65	0.052
Gemini-3.7-Flash	0.66	0.028	0.67	0.055	0.67	0.052
GLM-5.3-Flash	0.74	0.048	0.72	0.051	0.64	0.050
Deepseek-v4-flash	0.74	0.041	0.70	0.052	0.62	0.049

Appendix

Table E.2: Fit of the Bradley–Terry model. For each pair of rated sources, or of a rated source and the pooled opponent, the observed mean score is compared with the score the fitted model implies.

r

: correlation between the two across source pairs; MAR: mean absolute residual, weighted by the number of comparisons.

Figure F.1: Inversion rates for pairs in which exactly one item comes from a neutral source. Dot: one model; bar: median. Values in Table F.6 .

Rank	Source	Share	GPT	Luna	Gem	GLM	DS	Mav	Sct	L3.1	Tulu	Q3.5	Q3	Q2.5	Maj.
1	Amazon	12.2%	$-6$	$+16$	$+6$	$+15$	$+14$	$+27$	$+19$	$+6$	$+5$	$+26$	$+16$	$+15$	$+$
2	Walmart	6.5%	$+11$	$+24$	$+11$	$+17$	$+15$	$+25$	$+21$	$+12$	$+11$	$+20$	$+21$	$+17$	$+$
3	eBay	5.1%	$-18$	$+23$	$+12$	$+16$	$+9$	$-1$	$+3$	$+2$	$+0$	$+1$	$+3$	$+8$	–
4	Target	2.7%	$+5$	$+8$	$-0$	$+8$	$+5$	$+7$	$+3$	$+6$	$+12$	$+10$	$+8$	$+4$	$0$
5	YouTube	2.6%	$-29$	$-12$	$-22$	$-35$	$-35$	$-27$	$-14$	$-21$	$-22$	$-30$	$-22$	$-16$	$-$
6	Facebook	1.9%	$-31$	$-13$	$-21$	$-37$	$-35$	$-24$	$-19$	$-23$	$-19$	$-28$	$-23$	$-19$	$-$

Appendix

Table F.1: Shopping:

\tau_{s}

(percentage points) of the 20 sources that appear most often in the search results (the sources of Fig. 2 ), by model. Share: fraction of all search results, pooled over models. Cells are colored when the model labels the source Source

+

(red) or Source

-

(blue); gray: Source

0

. Maj.: the label held by a majority of the models that rate the source (–: none). Models: GPT = GPT-5.4-nano, Luna = GPT-5.6-Luna, Gem = Gemini-3.7-Flash, GLM = GLM-5.3-Flash, DS = Deepseek-v4-flash, Mav = Llama-4-Maverick, Sct = Llama-4-Scout, L3.1 = Llama-3.1-8B-It, Tulu = Tulu-3-8B, Q3.5 = Qwen3.5-27B, Q3 = Qwen3-30B-A3B, Q2.5 = Qwen2.5-32B.

Rank	Source	Share	GPT	Luna	Gem	GLM	DS	Mav	Sct	L3.1	Tulu	Q3.5	Q3	Q2.5	Maj.
1	Tripadvisor	8.9%	$-14$	$-12$	$-4$	$+5$	$-3$	$-11$	$-13$	$-8$	$-8$	$+2$	$-6$	$-10$	$-$
2	Facebook	6.5%	$-52$	$-42$	$-43$	$-51$	$-49$	$-43$	$-31$	$-19$	$-28$	$-46$	$-35$	$-32$	$-$
3	Expedia	4.3%	$-5$	$-9$	$-6$	$-7$	$-8$	$-21$	$-18$	$-3$	$-9$	$-18$	$-18$	$-8$	–
4	Booking.com	3.5%	$+11$	$+15$	$+6$	$+15$	$+13$	$+10$	$+7$	$+7$	$+9$	$+21$	$+8$	$+8$	$+$
5	Reddit	3.0%	$-51$	$-47$	$-39$	$-27$	$-25$	$-44$	$-35$	$-24$	$-31$	$-41$	$-35$	$-35$	$-$
6	Hotels.com	2.2%	$-7$	$-4$	$-6$	$+0$	$-3$	$-20$	$-17$	$-7$	$-7$	$-10$	$-20$	$-13$	–

Appendix

Table F.2: Accommodation: as Table F.1 .

Rank	Source	Share	GPT	Luna	Gem	GLM	DS	Mav	Sct	L3.1	Tulu	Q3.5	Q3	Q2.5	Maj.
1	arXiv	11.5%	$+34$	$+28$	$+29$	$+33$	$+27$	$+25$	$+28$	$+16$	$+19$	$+32$	$+23$	$+23$	$+$
2	ScienceDirect	4.2%	$+11$	$+18$	$+15$	$+17$	$+13$	$+2$	$-1$	$+3$	$+1$	$+14$	$+7$	$+8$	$+$
3	OpenReview	3.9%	$+29$	$+21$	$+21$	$+21$	$+14$	$+1$	$+3$	$+5$	$+3$	$+18$	$+3$	$+8$	$+$
4	Medium	3.9%	$-24$	$-23$	$-26$	$-21$	$-21$	$-11$	$-11$	$-5$	$-5$	$-24$	$-13$	$-7$	$-$
5	GitHub	3.6%	$-3$	$-7$	$-10$	$+14$	$+19$	$+0$	$+1$	$-6$	$-7$	$+12$	$+1$	$+4$	–
6	GitHub Pages	2.9%	$+17$	$+8$	$+8$	$+13$	$+11$	$+12$	$+13$	$+8$	$+8$	$+16$	$+11$	$+11$	$+$

Appendix

Table F.3: Scholar: as Table F.1 .

Domain	Source	Rank	Source $+$ (models, $\tau_{s}$ )	Source $-$ (models, $\tau_{s}$ )	Source $0$
Shopping	Amazon	1	11 models	GPT ( $-$ 6)	0
	eBay	3	Luna (+23), Gem (+12), GLM (+16), DS (+9), Q2.5 (+8)	GPT ( $-$ 18)	6
Accommodation	Forbes Travel Guide	34	Q3 (+22)	Mav ( $-$ 20)	10
Scholar	GitHub	5	GLM (+14), DS (+19), Q3.5 (+12), Q2.5 (+4)	Luna ( $-$ 7), Gem ( $-$ 10), L3.1 ( $-$ 6), Tulu ( $-$ 7)	4
	ResearchGate	8	Gem (+19), Q2.5 (+4)	GPT ( $-$ 11), Luna ( $-$ 5)	8
	Emergent Mind	12	Mav (+13), Sct (+6), L3.1 (+4), Tulu (+4), Q3 (+10), Q2.5 (+9)	GPT ( $-$ 17), Luna ( $-$ 22), Gem ( $-$ 15)	3

Appendix

Table F.4: Sources on which models disagree in direction: at least one model labels the source Source

+

and at least one labels it Source

-

(labels after FDR correction, § 3 ). Values in parentheses are

\tau_{s}

in percentage points; lists of seven or more models give only the count. Source

0

: number of models that rate the source but label it Source

0

. Rank: rank by frequency in the search results of the domain, pooled over models. Model abbreviations as in Table F.1 .

	Shopping			Accommodation			Scholar
Model	$\mathrm{Inv}_{-,+}$	$\mathrm{Inv}_{0,0}$	$\mathrm{Inv}_{+,-}$	$\mathrm{Inv}_{-,+}$	$\mathrm{Inv}_{0,0}$	$\mathrm{Inv}_{+,-}$	$\mathrm{Inv}_{-,+}$	$\mathrm{Inv}_{0,0}$	$\mathrm{Inv}_{+,-}$
GPT-5.4-nano	3 (2–5; 2,021)	20 (15–25; 1,646)	54 (46–62; 1,148)	7 (3–12; 1,226)	26 (21–31; 3,142)	72 (56–84; 380)	1 (1–1; 20,538)	19 (14–25; 1,770)	68 (64–73; 6,489)
GPT-5.6-Luna	0 (0–1; 7,313)	18 (7–33; 369)	92 (85–97; 644)	3 (1–6; 899)	14 (10–18; 5,310)	69 (52–83; 222)	2 (2–3; 24,847)	20 (16–25; 3,995)	69 (65–73; 10,597)
Gemini-3.7-Flash	0 (0–0; 926)	19 (7–34; 447)	97 (85–100; 39)	0 (0–0; 44)	10 (8–13; 8,288)	–	2 (1–3; 16,286)	9 (5–13; 1,428)	48 (42–53; 5,063)
GLM-5.3-Flash	1 (0–1; 3,147)	22 (16–29; 1,534)	71 (55–88; 334)	1 (0–2; 1,052)	16 (13–19; 7,783)	85 (70–96; 201)	1 (1–2; 12,336)	15 (11–18; 3,784)	63 (56–69; 2,737)
Deepseek-v4-flash	1 (0–1; 1,818)	18 (14–23; 1,378)	77 (63–89; 300)	2 (1–5; 790)	19 (15–23; 5,823)	50 (30–75; 133)	1 (1–2; 9,527)	17 (14–21; 3,545)	56 (49–64; 2,163)

Appendix

Table F.5: Inversion rates (%) of Fig. 3 . In parentheses: request-level bootstrap 95% confidence interval; number of comparisons with exactly one item selected. –: no pairs of this kind; Gemini-3.7-Flash labels only two Accommodation sources Source

+

, which yields no

(+,-)

pairs there. Across the three rates, the number of such comparisons per cell has a median of about 2,000 and is below 100 only for two Gemini-3.7-Flash cells.

	Shopping				Accommodation				Scholar
Model	$\mathrm{Inv}_{+,0}$	$\mathrm{Inv}_{0,-}$	$\mathrm{Inv}_{-,0}$	$\mathrm{Inv}_{0,+}$	$\mathrm{Inv}_{+,0}$	$\mathrm{Inv}_{0,-}$	$\mathrm{Inv}_{-,0}$	$\mathrm{Inv}_{0,+}$	$\mathrm{Inv}_{+,0}$	$\mathrm{Inv}_{0,-}$	$\mathrm{Inv}_{-,0}$	$\mathrm{Inv}_{0,+}$
GPT-5.4-nano	36 (28–45)	40 (33–46)	7 (5–10)	15 (11–20)	43 (28–59)	57 (51–64)	7 (5–9)	10 (5–17)	34 (29–38)	43 (35–52)	5 (3–8)	6 (5–7)
GPT-5.6-Luna	30 (19–42)	78 (56–91)	1 (0–2)	4 (1–7)	43 (30–55)	34 (27–42)	4 (3–6)	9 (4–16)	35 (31–39)	41 (34–47)	7 (5–10)	8 (6–9)
Gemini-3.7-Flash	27 (12–42)	81 (42–100)	1 (0–4)	8 (4–14)	57 (0–100)	59 (46–72)	1 (0–2)	2 (0–12)	21 (17–25)	26 (19–34)	5 (3–7)	4 (2–5)
GLM-5.3-Flash	32 (25–39)	84 (73–93)	1 (0–1)	9 (6–12)	38 (25–51)	58 (49–67)	3 (2–5)	5 (3–8)	34 (30–38)	35 (27–45)	4 (2–6)	6 (5–7)
Deepseek-v4-flash	32 (25–40)	74 (57–87)	3 (1–6)	12 (9–16)	34 (22–47)	48 (40–57)	5 (3–7)	6 (3–9)	31 (28–36)	33 (25–42)	4 (2–6)	7 (6–9)

Appendix

Table F.6: Inversion rates (%) of Eq. 2 for pairs in which exactly one item comes from a neutral source. In parentheses: request-level bootstrap 95% confidence interval.

Panel of Figure 21 — Figure G.1: Source-visibility replay prompt.

Model	WebShop	Accommodation	Scholar
GPT-5.4-nano	$59.59\%^{\ddagger}$ (446)	$62.03\%^{\ddagger}$ (158)	$59.32\%^{\ddagger}$ (1,846)
GPT-5.6-Luna	$63.48\%^{\ddagger}$ (397)	$60.06\%^{\ddagger}$ (154)	$57.04\%^{\ddagger}$ (1,979)
GLM-5.3-Flash	$62.90\%^{\ddagger}$ (248)	$56.37\%^{\ddagger}$ (106)	$55.20\%^{\ddagger}$ (990)
DeepSeek-v4-Flash	$63.23\%^{\ddagger}$ (189)	$55.91\%^{\ddagger}$ (110)	$54.83\%^{\ddagger}$ (890)
Gemini-3.7-Flash	$69.25\%^{\ddagger}$ (161)	$56.67\%$ (15)	$56.76\%^{\ddagger}$ (1,460)
Llama-4-Maverick	$82.35\%^{\ddagger}$ (289)	$57.46\%^{\ddagger}$ (248)	$60.70\%^{\ddagger}$ (1,138)

Appendix

Table H.1: Selection by displayed source identity with content fixed.

T_{+}

is the selection rate for the item displayed with a Source

+

source identity; parentheses give the number of complete, valid four-arm pairs. Two-sided tests of

H_{0}:T_{+}=50\%

use query-cluster standard errors and a Student

t

reference distribution.

\ast

\dagger

, and

\ddagger

denote unadjusted

p<0.05

p<0.01

, and

p<0.001

, respectively.

Figure I.1: Source-preference score

\tau_{s}

and mean request satisfaction among the 10, 20, and 30 most frequent sources within each model and domain. Each point represents one model–source observation. The three cutoffs yield 110, 220, and 330 observations per domain across 11 models, respectively. Mean satisfaction includes both selected and unselected items. Pearson’s

r

and Spearman’s

\rho

are computed over the plotted observations without averaging across models.

Panel of Figure 24 — Figure J.1: Prompt template used for synthetic-source DPO training, validation, and evaluation. Braced fields are substituted with the request and item information; the two source fields contain Kelto and Varnis in the assigned order. Line wrapping in the system message is for presentation only.

Setting	Value
Unique training pairs / rendered examples	5,000 / 10,000
Data-assignment and training seed	17
Optimizer updates	5,000
Microbatch size / accumulation steps	4 / 2
Effective batch size	8
Optimizer	Fused AdamW

Appendix

Table J.1: Shared settings for the fixed-data checkpoint-trajectory experiment. The training-set size counts unique content pairs; each pair yields two examples after reversing item order.

Panel of Figure 26 — Figure K.1: Reasons that the selected retailer will offer a lower price.

Panel of Figure 27 — Figure K.2: Reasons that the selected retailer will have price information available.

Explore similar work

In Agents We Trust, but Who Do Agents Trust? Latent Source Preferences Steer LLM Generations
Feb 17, 2026Mohammad Aflah Khan, Mahsa Amani, Soumi Das +5Language Model Bias Evaluation
Max Planck Institute for Software Systems · Microsoft, Hyderabad
Dear Algo: A Precision-First Agentic Intent Layer for Unified Search and Recommendation
Aug 16, 2026Rui Wang, Jiazhou Wang, Zheng Wei +16Language Model Steering Recommender Systems
Meta Platforms Inc. USA · Google DeepMind, USA
Does Rank Still Matter? Position Bias When AI Agents Shop on Our Behalf
Aug 24, 2026Davood Wadi, Yu MaAgentic Search Agentic Commerce

Feb 17, 2026J/K move · Enter open · S save

In Agents We Trust, but Who Do Agents Trust? Latent Source Preferences Steer LLM Generations

Mohammad Aflah Khan, Mahsa Amani, Soumi Das, Bishwamittra Ghosh+4

Max Planck Institute for Software Systems · Microsoft, Hyderabad

Language Model Bias Evaluation

Open paper AskPDF arXiv

Agents based on Large Language Models (LLMs) are increasingly being deployed as interfaces to information on online platforms. These agents filter, prioritize, and synthesize information retrieved from the platforms' back-end databases or via web search. In these scenarios, LLM agents govern the information users receive, by drawing users' attention to particular instances of retrieved information at the expense of others. While much prior work has focused on biases in the information LLMs themselves generate, less attention has been paid to the factors that influence what information LLMs select and present to users. We hypothesize that when information is attributed to specific sources (e.g., particular publishers, journals, or platforms), current LLMs exhibit systematic latent source preferences- that is, they prioritize information from some sources over others. Through controlled experiments on twelve LLMs from six model providers, spanning both synthetic and real-world tasks, we find that several models consistently exhibit strong and predictable source preferences. These preferences are sensitive to contextual framing, can outweigh the influence of content itself, and persist despite explicit prompting to avoid them. They also help explain phenomena such as the observed left-leaning skew in news recommendations in prior work. Our findings advocate for deeper investigation into the origins of these preferences, as well as for mechanisms that provide users with transparency and control over the biases guiding LLM-powered agents.

Abstract

Explore similar work

In Agents We Trust, but Who Do Agents Trust? Latent Source Preferences Steer LLM Generations

Dear Algo: A Precision-First Agentic Intent Layer for Unified Search and Recommendation

Does Rank Still Matter? Position Bias When AI Agents Shop on Our Behalf